re PR c/89946 (ICE in assemble_start_function, at varasm.c:1871)

PR c/89946
* varasm.c (assemble_start_function): Don't use tree_fits_uhwi_p
and gcc_unreachable if it fails, just call tree_to_uhwi which
verifies that too.  Test TREE_CHAIN instead of list_length > 1.
Start warning message with a lower-case letter.  Formatting fixes.
c-family/
* c-attribs.c (handle_patchable_function_entry_attribute): Add
function comment.  Warn if arguments of the attribute are not positive
integer constants.
testsuite/
* c-c++-common/pr89946.c: New test.

From-SVN: r270305

re PR rtl-optimization/90026 (ICE: verify_flow_info failed (error: missing barrier after block 2))

PR rtl-optimization/90026
* cfgcleanup.c (try_optimize_cfg): When removing empty bb with no
successors, look for BARRIERs inside of the whole BB_FOOTER chain
rather than just at the start of it.  If e->src BB_FOOTER is not NULL
in cfglayout mode, use emit_barrier_after_bb.

* g++.dg/opt/pr90026.C: New test.

From-SVN: r270304

libphobos: Parallelize the libphobos testsuite

The removal of the check programs in libphobos means that all tests
will now be compiled one at a time, causing a notable slowdown in
comparison with the previous method of testing.

libphobos/ChangeLog:

2019-04-12  Iain Buclaw  <ibuclaw@gdcproject.org>

* testsuite/Makefile.am:
* testsuite/Makefile.am (AUTOMAKE_OPTIONS): Remove dejagnu.
(RUNTEST): Remove variable.
(RUNTESTDEFAULTFLAGS, check_p_subno, check_p_numbers0,
check_p_numbers1, check_p_numbers2, check_p_numbers3,
check_p_numbers4, check_p_numbers5, check_p_numbers6, check_p_numbers,
check_p_subdirs, check_DEJAGNU_libphobos_targets): New variables.
(site.exp, %/site.exp, check-DEJAGNU, check-am, clean-local): New
rules written so that all the *.exp files are ran parallelized.
(CLEANFILES): Add *.exe, *.o, and site.exp.
* testsuite/Makefile.in: Regenerate.

From-SVN: r270303

libphobos: Replace library check programs with dg-runtest

Fixes multilib handling when running the libphobos testsuite.

libphobos/ChangeLog:

2019-04-12  Iain Buclaw  <ibuclaw@gdcproject.org>

PR d/89255
* configure.ac (ENABLE_SHARED, ENABLE_STATIC): Remove conditionals.
* configure: Regenerate.
* d_rules.am (%.t.lo, %.t.o): Remove rules.
(unittest_static_LINK, unittest_LINK, libgdruntime_t_la_LINK,
libgphobos_t_la_LINK): Remove variables.
* libdruntime/Makefile.am (DRUNTIME_TEST_LOBJECTS,
DRUNTIME_TEST_OBJECTS, check_PROGRAMS, check_LTLIBRARIES,
unittest_static_SOURCES, unittest_static_LIBTOOLFLAGS,
unittest_static_LDFLAGS, unittest_static_LDADD,
EXTRA_unittest_static_DEPENDENCIES, libgdruntime_t_la_SOURCES,
libgdruntime_t_la_LIBTOOLFLAGS, libgdruntime_t_la_LDFLAGS,
libgdruntime_t_la_LIBADD, EXTRA_libgdruntime_t_la_DEPENDENCIES,
unittest_SOURCES, unittest_LIBTOOLFLAGS, unittest_LDFLAGS,
unittest_LDADD): Remove variables.
(clean-local): Remove rule.
* libdruntime/Makefile.in: Regenerate.
* src/Makefile.am (PHOBOS_TEST_LOBJECTS,
PHOBOS_TEST_OBJECTS, check_PROGRAMS, check_LTLIBRARIES,
unittest_static_SOURCES, unittest_static_LIBTOOLFLAGS,
unittest_static_LDFLAGS, unittest_static_LDADD,
EXTRA_unittest_static_DEPENDENCIES, libgdruntime_t_la_SOURCES,
libgdruntime_t_la_LIBTOOLFLAGS, libgdruntime_t_la_LDFLAGS,
libgdruntime_t_la_LIBADD, EXTRA_libgdruntime_t_la_DEPENDENCIES,
unittest_SOURCES, unittest_LIBTOOLFLAGS, unittest_LDFLAGS,
unittest_LDADD): Remove variables.
(clean-local): Remove rule.
* src/Makefile.in: Regenerate.
* testsuite/lib/libphobos.exp (libphobos_test_name): New global.
(libphobos_init): Set libphobos_test_name.
(libphobos-dg-test): Override name if libphobos_test_name is set.
(filter_libphobos_unittests): New proc.
* testsuite/libphobos.aa/aa.exp: Don't load libphobos-dg.exp.
* testsuite/libphobos.cycles/cycles.exp: Set libphobos_test_name.
* testsuite/libphobos.druntime/druntime.exp: New file.
* testsuite/libphobos.druntime_shared/druntime_shared.exp: New file.
* testsuite/libphobos.phobos/phobos.exp: New file.
* testsuite/libphobos.phobos_shared/phobos_shared.exp: New file.
* testsuite/libphobos.shared/shared.exp: Use dg-runtest to run tests.
* testsuite/libphobos.hash/hash.exp: Don't load libphobos-dg.exp
* testsuite/libphobos.init_fini/init_fini.exp: Likewise.
* testsuite/libphobos.thread/thread.exp: Likewise.
* testsuite/libphobos.typeinfo/typeinfo.exp: Likewise.
* testsuite/libphobos.unittests/unittests.exp: Remove.
* testsuite/test_runner.d: Remove.
* testsuite/testsuite_flags.in: Add phobos to --gdcpaths.

From-SVN: r270302

d: Add -fbuilding-libphobos-tests option

Currently, the druntime and phobos unittests are compiled as a separate
check program, then ran by the libphobos.unittest/unittest.exp script.
As PR d/89255 notes, this process lacks proper multilib handling.

As a first step, a new internal option that instructs the compiler to
put the reference to all unittest functions in another symbol has been
added.  This will allow each module to be compiled separately as a
standalone program using dg-runtest without running into collisions in
the D runtime module registry.

gcc/d/ChangeLog:

2019-04-12  Iain Buclaw  <ibuclaw@gdcproject.org>

* d-tree.h (DECL_IN_UNITTEST_CONDITION_P): Define.
* decl.cc (DeclVisitor): Add in_version_unittest_ field.
(DeclVisitor::visit(ConditionalDeclaration)): New override.
(DeclVisitor::visit(FuncDeclaration)): Set
DECL_IN_UNITTEST_CONDITION_P.
* lang.opt (-fbuilding-libphobos-tests): Add option.
* modules.cc (current_testing_module): New static variable.
(build_module_tree): Generate second moduleinfo symbol to hold
reference to unittests if flag_building_libphobos_tests.
(register_module_decl): Check DECL_IN_UNITTEST_CONDITION_P to decide
which moduleinfo the decl should be registered against.

From-SVN: r270301

d/dmd: Merge upstream dmd c185f9df1

Adds new virtual isVersionCondition, this is so that in the code
generation pass, a ConditionDeclaration's condition can be identified
without requiring a Visitor function.

Reviewed-on: https://github.com/dlang/dmd/pull/9591

From-SVN: r270300

Daily bump.

From-SVN: r270299

libphobos: Merge upstream phobos cf95639ff

Backports ports committed to upstream phobos since the last sync.

Reviewed-on: https://github.com/dlang/phobos/pull/6947

From-SVN: r270296

libphobos: Merge upstream druntime 175bf5fc

Backports extern(C) bindings committed to upstream druntime since the
last sync.

Reviewed-on: https://github.com/dlang/druntime/pull/2552

From-SVN: r270295

d/dmd: Merge upstream dmd d7ed327ed

Fixes ICE when accessing empty array in CTFE.

Reviewed-on: https://github.com/dlang/dmd/pull/9563

From-SVN: r270294

Replace direct PSTL uses of assert() with a macro

This also replaces calls to __TBB_ASSERT so that there are two macro
definitions provided by c++config -
    __PSTL_ASSERT(_Condition)
    __PSTL_ASSERT_MSG(_Condition, _Message)

    * include/bits/c++config:
    Add definition for __PSTL_ASSERT.
    Add definition for __PSTL_ASSERT_MSG.
    * include/pstl/algorithm_impl.h: Replace use of assert().
    * include/pstl/numeric_impl.h: Replace use of assert().
    * include/pstl/parallel_backend_tbb.h:
    Replace use of assert().
    Replace use of __TBB_ASSERT().

    * include/pstl/parallel_backend_utils.h: Replace use of assert().

From-SVN: r270293

re PR translation/89939 (messages for translation must not contain embedded macro parameters)

2019-04-11  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR translation/89939
* frontend-passes.c (B_ERROR): Delete macro.
(C_ERROR): Delete macro.
(B_ERROR_1): New macro.
(C_ERROR_1): New macro.
(C_ERROR_2): New macro.
(inline_matmul_assign): Use new macros.
(call_external_blas): Likewise.

From-SVN: r270292

PR libstdc++/90046 fix build failure on epiphany-elf

The epiphany-elf target aligns structs to 8 bytes, which causes the
static_assert(alignof(_Chunk) == 1) to fail.

Instead of requiring _Chunks to be positionable at any alignment, ensure
new buffers are aligned to alignof(_Chunk). Because the buffer size is a
power of two, we know that both the buffer size and sizeof(_Chunk) are
multiples of alignof(_Chunk). So is p is aligned to alignof(_Chunk) then
so is (p + size - sizeof(_Chunk)). So just ensure the new buffer is
aligned to at least alignof(_Chunk), which should already be true
because the caller requests at least alignof(max_align_t).

PR libstdc++/90046
* src/c++17/memory_resource.cc
(monotonic_buffer_resource::_Chunk::allocate): Increase alignment if
needed to allow placing a _Chunk at the end of the buffer.
(monotonic_buffer_resource::_M_new_buffer): Remove static_assert.

From-SVN: r270291

re PR rtl-optimization/87763 (aarch64 target testcases fail after r265398)

2018-04-11  Steve Ellcey  <sellcey@marvell.com>

PR rtl-optimization/87763
* gcc.target/aarch64/combine_bfxil.c: Change some bfxil checks
to bfi.

From-SVN: r270289

re PR rtl-optimization/87763 (aarch64 target testcases fail after r265398)

2018-04-11  Steve Ellcey  <sellcey@marvell.com>

PR rtl-optimization/87763
* config/aarch64/aarch64.md (*aarch64_bfi<GPI:mode>4_noshift_alt):
New Instruction.

From-SVN: r270288

[doc] Note variable shadowing at max macro using statement expression

When suggesting to rewrite the unsafe (with respect to multiple evaluation of
arguments) macro definition:
...
  #define max(a,b) ((a) > (b) ? (a) : (b))
...
into the safe macro definition:
...
  #define maxint(a,b) \
    ({int _a = (a), _b = (b); _a > _b ? _a : _b; })
...
mention the variable shadowing problem for:
...
  #define maxint3(a, b, c) \
    ({int _a = (a), _b = (b), _c = (c); maxint (maxint (_a, _b), _c); })
...

2019-04-11  Tom de Vries  <tdevries@suse.de>

* doc/extend.texi (@node Statement Exprs): Note variable shadowing at
max macro using statement expression.

From-SVN: r270287

re PR translation/90035 (Non-translatable C++ parser diagnostics)

PR translation/90035
* parser.h (struct cp_parser): Add
type_definition_forbidden_message_arg member.
* parser.c (cp_debug_parser): Print it.
(cp_parser_check_type_definition): Pass
parser->type_definition_forbidden_message_arg as second argument to
error.
(cp_parser_has_attribute_expression, cp_parser_sizeof_operand): Set
parser->type_definition_forbidden_message_arg and use G_() with
%qs for parser->type_definition_forbidden_message instead of
building untranslatable message using concat.

From-SVN: r270286

xcoffout.h (xcoff_private_rodata_section_name): Declare.

        * xcoffout.h (xcoff_private_rodata_section_name): Declare.
        * xcoffout.c (xcoff_private_rodata_section_name): Define.
        * config/rs6000/rs6000.c (rs6000_xcoff_asm_init_sections): Create
        read_only_private_data_section using xcoff_private_rodata_section_name.
        (rs6000_xcoff_file_start): Generate xcoff_private_rodata_section_name.

From-SVN: r270284

[aarch64] PR90016 - aarch64: reference to undeclared N in help for command line option

2019-04-11  Christophe Lyon  <christophe.lyon@linaro.org>

PR target/90016
* config/aarch64/aarch64.opt (msve-vector-bits): Add missing final '.'.

From-SVN: r270280

re PR rtl-optimization/89965 (wrong code with -O -mtune=nano-x2 -fcaller-saves -fexpensive-optimizations -fno-tree-dce -fno-tree-ter)

PR rtl-optimization/89965
* dce.c (sp_based_mem_offset): New function.
(find_call_stack_args): Use sp_based_mem_offset.

From-SVN: r270278

Clarify documentation for -flive-patching

* doc/invoke.texi (Optimize Options): Clarify -flive-patching docs.

From-SVN: r270276

re PR tree-optimization/90020 (-O2 -Os x86-64 wrong code generated for GNU Emacs)

2019-04-11  Richard Biener  <rguenther@suse.de>

PR tree-optimization/90020
* tree-ssa-sccvn.c (vn_reference_may_trap): New function.
* tree-ssa-sccvn.h (vn_reference_may_trap): Declare.
* tree-ssa-pre.c (compute_avail): Use it to not put
possibly trapping references after a call that might not
return into EXP_GEN.
* gcse.c (compute_hash_table_work): Do not elide
marking a block containing a call if the call might not
return.

* gcc.dg/torture/pr90020.c: New testcase.

From-SVN: r270275

re PR tree-optimization/90018 (r265453 miscompiled 527.cam4_r in SPEC CPU 2017)

2019-04-11  Richard Biener  <rguenther@suse.de>

PR tree-optimization/90018
* tree-vect-data-refs.c (vect_preserves_scalar_order_p):
Test both SLP and interleaving variants.

* gcc.dg/vect/pr90018.c: New testcase.

From-SVN: r270273

S/390: Add arch13 pipeline and model number

This patch adds the pipeline description and the CPU model number for
arch13.

From-SVN: r270272

Daily bump.

From-SVN: r270271

re PR rtl-optimization/87763 (aarch64 target testcases fail after r265398)

2018-04-10  Steve Ellcey  <sellcey@marvell.com>

PR rtl-optimization/87763
* gcc.target/aarch64/combine_bfxil.c: Change some bfxil checks
to bfi.
* gcc.target/aarch64/combine_bfi_2.c: New test.

From-SVN: r270267

re PR rtl-optimization/87763 (aarch64 target testcases fail after r265398)

2018-04-10  Steve Ellcey  <sellcey@marvell.com>

PR rtl-optimization/87763
* config/aarch64/aarch64-protos.h (aarch64_masks_and_shift_for_bfi_p):
New prototype.
* config/aarch64/aarch64.c (aarch64_masks_and_shift_for_bfi_p):
New function.
* config/aarch64/aarch64.md (*aarch64_bfi<GPI:mode>5_shift):
New instruction.
(*aarch64_bfi<GPI:mode>5_shift_alt): Ditto.
(*aarch64_bfi<GPI:mode>4_noand): Ditto.
(*aarch64_bfi<GPI:mode>4_noand_alt): Ditto.
(*aarch64_bfi<GPI:mode>4_noshift): Ditto.

From-SVN: r270266

Update documentation regarding bogus memory leaks in libstdc++

* doc/xml/faq.xml: Add information about emergency EH pool.
* doc/xml/manual/debug.xml: Update list of memory debugging tools.
Move outdated information on mt_allocator to a separate section.
* doc/xml/manual/evolution.xml: Clarify that GLIBCXX_FORCE_NEW
doesn't affect the default allocator.

From-SVN: r270264

Change wording of -fipa-icf documentation

* doc/invoke.texi (Optimize Options): Change "Nevertheless" to
"Although" in -fipa-icf documentation.

From-SVN: r270260

Fix typo in effective-target check

* testsuite/lib/libstdc++.exp (check_v3_target_parallel_mode): Fix
typo.

From-SVN: r270259

Fix a stack exhaustion bug in libiberty's demangler when decoding a pathalogically constructed mangled name.

PR 89394
* cp-demangle.c (cplus_demangle_fill_name): Reject negative
lengths.
(d_count_templates_scopes): Replace num_templates and num_scopes
parameters with a struct d_print_info pointer parameter.  Adjust
body of the function accordingly.  Add recursion counter and check
that the recursion limit is not reached.
(d_print_init): Pass dpi parameter to d_count_templates_scopes.
Reset recursion counter afterwards, unless the recursion limit was
reached.

From-SVN: r270258

Clearly document behaviour of multiple -g options

This copies the wording from the -O options to clearly state what
happens if more than one -g option is used.

* doc/invoke.texi (Debugging Options): Explicitly state the semantics
of using multiple -g options.

From-SVN: r270257

Make gcov docs more precise (PR gcov-profile/89959).

2019-04-10  Martin Liska  <mliska@suse.cz>

PR gcov-profile/89959
* doc/gcov.texi: Make documentation of -x option
more precise.

From-SVN: r270251

tree-vectorizer.h (_stmt_vec_info): Remove same_dr_stmt member.

2019-04-10  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (_stmt_vec_info): Remove same_dr_stmt
member.
(DR_GROUP_SAME_DR_STMT): Remove.
* tree-vect-stmts.c (vectorizable_load): Remove unreachable code.
* tree-vect-data-refs.c (vect_analyze_group_access_1): Likewise,
replace with assert.
(vect_analyze_data_ref_accesses): Fix INTEGER_CST comparison.
(vect_record_grouped_load_vectors): Remove unreachable code.

From-SVN: r270250

PR libstdc++/89851 Add testcase for std::variant equality

Add a test for the regression introduced with r269422 and fixed with
r270056.

PR libstdc++/89851
* testsuite/20_util/variant/89851.cc: New test.

From-SVN: r270249

[aarch64] PR90016 - aarch64: reference to undeclared N in help for command line option

'to N' is now redundant and misleading given the earlier change to use
<number>.

Removed.

PR target/90016
* config/aarch64/aarch64.opt (msve-vector-bits): Remove redundant and
obsolete reference to N.

From-SVN: r270248

re PR middle-end/90025 (botan2 miscompilation on s390x-linux since r268957)

PR middle-end/90025
* expr.c (store_expr): Set properly size on the MEM passed to
clear_storage.

* gcc.c-torture/execute/pr90025.c: New test.

From-SVN: r270247

re PR c++/90010 (valgrind error with snprintf and -Wall)

PR c++/90010
* gimple-ssa-sprintf.c (target_to_host): Fix handling of targstr
with strlen in between hostsz-3 and hostsz-1 inclusive when no
translation is needed, and when translation is needed, only append
... if the string length is hostsz or more bytes long.  Avoid using
strncpy or strcat.

* gcc.dg/pr90010.c: New test.

From-SVN: r270246

Daily bump.

From-SVN: r270245

c-ada-spec.c (print_destructor): Deal with deleting destructors.

* c-ada-spec.c (print_destructor): Deal with deleting destructors.
(dump_ada_declaration) <FUNCTION_DECL>: Likewise.

From-SVN: r270239

Add comments and style fixes to <variant>

* include/std/variant: Adjust whitespace. Add comments.
(_Multi_array): Leave primary template undefined.
(_Multi_array<_Tp>): Define partial specialization for base case of
recursion.
(__gen_vtable_impl, __gen_vtable): Remove redundant && from type
which is always a reference.
(__gen_vtable::_S_apply()): Remove function, inline body into
default member initializer.
* testsuite/20_util/variant/visit.cc: Test with noncopyable types.

From-SVN: r270238

Fix std::visit to support arbitrary callables

The __visitor_result_type helper didn't use std::invoke and so didn't
compile when the visitor was a pointer-to-member rather than a function
object. Use std::invoke_result instead.

* include/std/variant (__variant_idx_cookie): Add member type.
(__visitor_result_type): Remove.
(__do_visit): Use invoke_result instead of __visitor_result_type.
* testsuite/20_util/variant/visit.cc: New test.

From-SVN: r270237

PR libstdc++/90008 remove unused capture from variant rel ops

PR libstdc++/90008
* include/std/variant (_VARIANT_RELATION_FUNCTION_TEMPLATE): Remove
unused capture.
* testsuite/20_util/variant/90008.cc: New test.

From-SVN: r270236

* gcc.target/i386/ifcvt-onecmpl-abs-1.c
(dg-options): Use -O2 -fdump-rtl-ce1.
(dg-final): Scan ce1 RTL dump instead of asm dump.

From-SVN: r270234

Add PSTL internal namespace qualifications

* include/pstl/algorithm_impl.h: Add namespace qualification.
* include/pstl/execution_defs.h: Add namespace qualification.
* include/pstl/execution_impl.h: Add namespace qualification.
* include/pstl/numeric_impl.h: Add namespace qualification.
* include/pstl/parallel_backend_tbb.h: Add namespace qualification.
* include/pstl/unseq_backend_simd.h: Add namespace qualification.
* include/pstl/parallel_backend_utils.h: Include <cassert>.

From-SVN: r270231

re PR translation/90011 (trailing space in diagnostic)

PR translation/90011
* typeck2.c (check_narrowing): Remove trailing space from diagnostics.

From-SVN: r270229

Hi there,

The "*neon_mov<mode>" patterns for 128 bit sized quantities uses the "Dn"
constraint to match vmov.f32 and vmov.i<vec-width> patterns.
This constraint boils down to using the `neon_immediate_valid` function.
Once the constraint has matched, the output C statement asserts that function
passes.

The output C statement calls `neon_immediate_valid` with the mode taken from the
iterator, while the constraint takes the mode from the operand.
This can cause a discrepency when the operand is a CONST_INT, as the constraint
passes VOIDmode which `neon_immediate_valid` treats as DImode, while the C
statement passes the mode of the iterator which can be TImode.
When this happens, the `neon_immediate_valid` can fail in the second call (if
e.g. the CONST_INT is a valid immediate in DImode but not TImode) which would
trigger the assertion.

The testcase added with this patch triggers this when compiled with an arm cross
compiler using the command line below.
gcc -march=armv8-a -c neon-immediate-timode.c -O1 -mfloat-abi=hard -mfpu=neon-fp-armv8

This patch splits the original "Dn" constraint into three new constraints, "DN"
for TImode CONST_INT, "Dn" for DImode CONST_INT, and "Dm" for CONST_VECTOR.
Splitting things up this way requires using one extra alternative in the
"*neon_mov<mode>" patterns, but makes it clear from the constraint what mode is
being used.

We also remove the behaviour of treating VOIDmode as DImode in
`neon_valid_immediate` since the original "Dn" constraint was the only place
that functionality was used.  VOIDmode is now never passed to that function.
An assertion has been added to the function to ensure this problem is caught
earlier on.

Bootstrapped on arm-none-linux-gnueabihf
Regtested on cross-compiler arm-none-eabi

gcc/ChangeLog:

2019-04-09  Matthew Malcomson  <matthew.malcomson@arm.com>

PR target/90024
* config/arm/arm.c (neon_valid_immediate): Disallow VOIDmode parameter.
* config/arm/constraints.md (Dm, DN, Dn): Split previous Dn constraint
into three.
* config/arm/neon.md (*neon_mov<mode>): Account for TImode and DImode
differences directly.
(*smax<mode>3_neon, vashl<mode>3, vashr<mode>3_imm): Use Dm constraint.

gcc/testsuite/ChangeLog:

2019-04-09  Matthew Malcomson  <matthew.malcomson@arm.com>

PR target/90024
* gcc.dg/torture/neon-immediate-timode.c: New test.

From-SVN: r270226

re PR translation/90011 (trailing space in diagnostic)

PR translation/90011
* ipa-devirt.c (compare_virtual_tables): Remove two trailing spaces
from diagnostics.
* config/arm/freebsd.h (LINK_SPEC): Remove trailing space from -p
diagnostics.
* config/riscv/freebsd.h (LINK_SPEC): Likewise.
* config/aarch64/aarch64-freebsd.h (FBSD_TARGET_LINK_SPEC): Likewise.
* config/darwin.h (DRIVER_SELF_SPECS, ASM_FINAL_SPEC): Remove
trailing space from -gsplit-dwarf diagnostics.

From-SVN: r270225

re PR middle-end/89998 (ICE: verify_gimple failed in printf-return-value)

PR tree-optimization/89998
* gimple-ssa-sprintf.c (try_substitute_return_value): Use lhs type
instead of integer_type_node if possible, don't add ranges if return
type is not compatible with int.
* gimple-fold.c (gimple_fold_builtin_sprintf,
gimple_fold_builtin_snprintf): Use lhs type instead of hardcoded
integer_type_node.

* gcc.c-torture/compile/pr89998-1.c: New test.
* gcc.c-torture/compile/pr89998-2.c: New test.

From-SVN: r270224

Come up with bootstrap-lto-lean config.

2019-04-09  Martin Liska  <mliska@suse.cz>

* Makefile.in: Regenerate.
* Makefile.tpl: Pass GENERATOR_CFLAGS
in all stages.
2019-04-09  Martin Liska  <mliska@suse.cz>

* bootstrap-lto-lean.mk: New file.
2019-04-09  Martin Liska  <mliska@suse.cz>

* Makefile.in: Use GENERATOR_CFLAGS for all generators.
* doc/install.texi: Document the new config.

From-SVN: r270223

SVE fallout from PR90006

2019-04-08  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-data-refs.c (vect_get_smallest_scalar_type): Always
use gimple_expr_type for load and store calls.  Skip over the
condition argument in a conditional internal function.
Protect use of TREE_INT_CST_LOW.

From-SVN: r270222

re PR target/90015 (riscv: typo "intterupt" in diagnostic)

PR target/90015
* config/riscv/riscv.c (riscv_get_interrupt_type): Fix comment typo.
(riscv_merge_decl_attributes): Fix typo in diagnostics.  Remove
trailing period from it too.

* gcc.target/riscv/interrupt-conflict-mode.c (foo): Adjust expected
diagnostics.

From-SVN: r270221

compiler: sort packages in export data more deterministically

    We can have multiple packages with the same name, so also sort by pkgpath.
    To avoid an inconsistent sort, sort by symbol and pointer address if
    we somehow get two different packages with the same name and pkgpath.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/171032

From-SVN: r270220

Daily bump.

From-SVN: r270219

Fix visit<R> for variant.

* include/std/variant (__do_visit): Add a template parameter
for enforcing same return types for visit.
(__gen_vtable_impl): Likewise.
(_S_apply_single_alt): Adjust.
(__visit_invoke_impl): New. Handle casting to void.
(__do_visit_invoke): New. Enforces same return types.
(__do_visit_invoke_r): New. Converts return types.
(__visit_invoke): Adjust.
(__gen_vtable):  Add a template parameter for enforcing
same return types for visit.
* testsuite/20_util/variant/visit_r.cc: Add a test for a visitor with
different return types.
* testsuite/20_util/variant/visit_neg.cc: New. Ensures that
visitors with different return types don't accidentally
compile with regular visitation.

From-SVN: r270216

typeck2.c (digest_init_r): Don't condition the object slicing warning on flag_checking.

* typeck2.c (digest_init_r): Don't condition the object slicing warning
on flag_checking.

From-SVN: r270215

libgo: update to Go 1.12.2

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/170706

From-SVN: r270214

[AArch64] Add tsv110 pipeline scheduling

Committed on behalf of Wu Yuan:

2019-04-04  wu yuan <wuyuan5@huawei.com>

* config/aarch64/aarch64-cores.def (tsv1100): Change scheduling model.
* config/aarch64/aarch64.md : Add "tsv110.md"
* config/aarch64/tsv110.md: New file.

From-SVN: r270212

re PR tree-optimization/90006 (gcc loops indefinitely around vect_get_constant_vectors on -O2 -ftree-slp-vectorize -fno-math-errno)

2019-04-08  Richard Biener  <rguenther@suse.de>

PR tree-optimization/90006
* tree-vect-data-refs.c (vect_get_smallest_scalar_type): Handle
calls like lrint.

* gcc.dg/vect/bb-slp-pr90006.c: New testcase.

From-SVN: r270210

LOCAL_PATCHES: Add revision.

From-SVN: r270209

Add missing libsanitizer extra patch (r259664) (PR sanitizer/89941).

2019-04-08  Martin Liska  <mliska@suse.cz>

PR sanitizer/89941
* sanitizer_common/sanitizer_platform_limits_linux.cc (defined):
Reapply patch from r259664.
* sanitizer_common/sanitizer_platform_limits_posix.h (defined):
Likewise.

From-SVN: r270208

The fma_forest, fma_root_node and func_fma_steering classes lack a copy constructor.

The fma_forest, fma_root_node and func_fma_steering classes lack a
copy constructor.  However, they contain pointers to allocated memory
so this omission can be regarded as poor style.  We don't need to copy
such objects, so declare the copy constructor private to inhibit
accidental copying.

2019-04-08  Andrea Corallo  <andrea.corallo@arm.com>

PR target/83033
* config/aarch64/cortex-a57-fma-steering.c (fma_forest): Prohibit copy
construction.
(fma_root_node): Likewise.
(func_fma_steering): Likewise.

From-SVN: r270207

re PR rtl-optimization/89865 (FAIL: gcc.target/i386/pr49095.c scan-assembler-times \\\\), % 45)

PR rtl-optimization/89865
* config/i386/i386.md: Add peepholes for z = x; x ^= y; x != z.

* gcc.target/i386/pr49095.c: Don't expect any RMW sequences.

From-SVN: r270206

re PR rtl-optimization/89865 (FAIL: gcc.target/i386/pr49095.c scan-assembler-times \\\\), % 45)

PR rtl-optimization/89865
* config/i386/i386.md
(SWI12 peephole for mem {+,-,&,|,^}= x; mem != 0): Fix up operand
numbers not to clash with the additional operands[4].
(peepholes for mem {+,-,&,|,^}= x; mem != 0): New peephole2s
with extra register copy in the middle.

* gcc.target/i386/pr49095.c: Adjust number of expected RMW spots
on ia32.

From-SVN: r270205

Add data_file to GCOV interm. format (PR gcov-profile/89961).

2019-04-08  Martin Liska  <mliska@suse.cz>

PR gcov-profile/89961
* doc/gcov.texi: Document data_file.
* gcov.c (generate_results): Add data_info into JSON output.

From-SVN: r270204

re PR middle-end/89725 (ICE in get_fnname_from_decl, at varasm.c:1723)

2019-04-01  Bin Cheng  <bin.cheng@linux.alibaba.com>

PR tree-optimization/89725
* tree-chrec.c (chrec_contains_symbols): New parameter.  Handle outer
loop's chrec as invariant symbol.
* tree-chrec.h (chrec_contains_symbols): New parameter.
* tree-data-ref.c (analyze_miv_subscript): Pass new argument.
(build_classic_dist_vector_1, add_other_self_distances): Bypass access
function of loops not in DDR's loop_nest.
* tree-data-ref.h (index_in_loop_nest): Add unreachable check.

* gcc.dg/tree-ssa/pr89725.c: New test.

From-SVN: r270203

Fix expected scanned pattern.

2019-04-08  Martin Liska  <mliska@suse.cz>

* gcc.target/riscv/arch-1.c: Fix expected scanned pattern.

From-SVN: r270202

re PR c++/89914 (ICE in nothrow_spec_p, at cp/except.c:1238)

/cp
2019-04-08  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/89914
* semantics.c (trait_expr_value): Don't use TYPE_NOTHROW_P
when maybe_instantiate_noexcept fails.
(classtype_has_nothrow_assign_or_copy_p): Likewise.
* method.c (implicitly_declare_fn): Avoid passing error_mark_node
to build_exception_variant.

/testsuite
2019-04-08  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/89914
* g++.dg/ext/has_nothrow_constructor-3.C: New.

From-SVN: r270201

[testsuite,libstdc++-v3] Fix 27_io/filesystem/iterators/caching.cc

2019-04-08  Christophe Lyon  <christophe.lyon@linaro.org>

* testsuite/27_io/filesystem/iterators/caching.cc: Add
dg-require-filesystem-ts.

From-SVN: r270199

re PR target/89623 (Can't build mips-wrs-vxworks cross-compiler)

2019-04-08  Chenghua Xu  <paul.hua.gm@gmail.com>

gcc/
PR target/89623
* config/mips/mips.opt (LOONGSON_EXT2): Use Var instead of
Mask.

From-SVN: r270198

Daily bump.

From-SVN: r270195

re PR fortran/89904 (ICE in gfortran starting with r270045)

2019-04-07  Harald Anlauf  <anlauf@gmx.de>

PR fortran/89904
* ChangeLog: Correct ChangeLog entry.

From-SVN: r270192

ChangeLog: Fix whitespace.

* ChangeLog: Fix whitespace.
* testsuite/ChangeLog: Ditto.

From-SVN: r270191

re PR target/89945 (ICE in gen_lowpart_general, at rtlhooks.c:63)

PR target/89945
* config/i386/i386.md (anddi_1 to andsi_1_zext splitter):
Avoid calling gen_lowpart with SYMBOL_REF and LABEL_REF operand.

testsuite/ChangeLog:

PR target/89945
* gcc.target/i386/pr89945.c: New test.

From-SVN: r270190

Regenerate libiberty/functions.texi

This patch just picks up changes made in r231983 (git 2afaef0360)
and r268793 (git 0f7c25df5f).

* functions.texi: Regenerate.

From-SVN: r270189

c-ada-spec.c (is_float128): New predicate extracted from...

c-family/
* c-ada-spec.c (is_float128): New predicate extracted from...
(dump_ada_node) <COMPLEX_TYPE>: Use it to recognize __cfloat128.
<REAL_TYPE>: ...here.  Call it.
ada/
* libgnat/i-cexten.ads (CFloat_128): New type.

From-SVN: r270188

Daily bump.

From-SVN: r270187

re PR fortran/87352 (Large stack usage with new gfortran)

2019-04-06  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87352
* gfortran.h (gfc_component): Add finalized field.
* class.c (finalize_component): If the component is already
finalized, return early.  Set component->finalized on exit.

2019-04-06  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87352
* gfortran.dg/finalize_28.f90: Adjust count of __builtin_free.
* gfortran.dg/finalize_33.f90: Likewise.
* gfortran.dg/finalize_34.f90: New test.

From-SVN: r270184

re PR fortran/89981 (gfortran -pedantic rejects code in 8.3.1 that is accepted with 8.3.0)

2019-04-06  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/89981
* resolve.c (resolve_global_procedure): If the global symbol is an
ENTRY, also look up its name among the entries.

2019-04-06  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/89981
* gfortran.dg/entry_22.f90: New test.

From-SVN: r270182

Daily bump.

From-SVN: r270181

PR c++/87145 - bogus error converting class type in template arg list.

* pt.c (convert_nontype_argument): Don't call
build_converted_constant_expr if it could involve calling a conversion
function with a instantiation-dependent constructor as its argument.

* g++.dg/cpp0x/constexpr-conv3.C: New test.
* g++.dg/cpp0x/constexpr-conv4.C: New test.

From-SVN: r270178

PR bootstrap/89980 - pointer initialization with empty string folded to zero

gcc/cp/ChangeLog:

PR bootstrap/89980
* decl.c (reshape_init_array_1): Avoid treating empty strings
as zeros in array initializers.
Use trivial_type_p () instead of TYPE_HAS_TRIVIAL_DFLT().

gcc/testsuite/ChangeLog:

PR bootstrap/89980
* g++.dg/init/array52.C: New test.

From-SVN: r270177

Implement std::visit<R> for C++2a (P0655R1)

* doc/xml/manual/status_cxx2020.xml: Update status.
* include/std/variant (visit<R>): Define for C++2a (P0655R1).
* testsuite/20_util/variant/visit_r.cc: New test.

From-SVN: r270176

Use hidden friends for directory iterator comparisons

The equality operators for directory iterators are not explicitly
specified in the standard, they're only required to meet the iterator
requirements. This means we don't need to declare them at namespace
scope and can implement them as hidden friends.

Also add 'noexcept' to directory_iterator's dereference operators.

* include/bits/fs_dir.h (directory_iterator::operator*)
(directory_iterator::operator->): Add noexcept.
(operator==, operator!=): Replace namespace-scope equality operators
for directory iterators with hidden friends.

From-SVN: r270175

PR libstdc++/89986 export directory_iterator::increment

PR libstdc++/89986
* config/abi/pre/gnu.ver: Add missing exports.
* testsuite/27_io/filesystem/iterators/directory_iterator.cc: Test
increment member.

From-SVN: r270174

Share all recursive_directory_iterator state [LWG 2708]

Implement the proposed resolution of LWG 2708 by moving the _M_options
and _M_pending members out of the recursive_directory_iterator into the
shared _Dir_stack object. Because _Dir_stack is an opaque type, the
member functions that access the _M_options and _M_pending variables
cannot be inline. Move them into the library.

As a drive-by fix, add noexcept to the non-throwing member functions of
recursive_directory_iterator.

* config/abi/pre/gnu.ver: Export new symbols.
* include/bits/fs_dir.h (recursive_directory_iterator::options())
(recursive_directory_iterator::recursion_pending())
(recursive_directory_iterator::disable_recursion_pending()): Remove
inline definitions. Make noexcept.
(recursive_directory_iterator::depth())
(recursive_directory_iterator::operator*())
(recursive_directory_iterator::operator->()): Make noexcept.
(recursive_directory_iterator::_M_options)
(recursive_directory_iterator::_M_pending): Remove data members.
* src/c++17/fs_path.cc (_Dir_stack): Add constructor and data members.
(recursive_directory_iterator::recursive_directory_iterator): Remove
ctor-initializer. Use new constructor for _Dir_stack.
(recursive_directory_iterator::options())
(recursive_directory_iterator::recursion_pending())
(recursive_directory_iterator::disable_recursion_pending()): Add
non-inline definitions.
(recursive_directory_iterator::depth()): Make noexcept.
(recursive_directory_iterator::increment(error_code&))
(recursive_directory_iterator::pop(error_code&)): Adjust to new
location of options and recursion_pending members.
* testsuite/27_io/filesystem/iterators/recursion_pending.cc: New test.
* testsuite/util/testsuite_fs.h (__gnu_test::scoped_file): Add
user-declared move constructor and assignment operator, to make the
type move-only.

From-SVN: r270173

Fix directory_iterator handling of DT_UNKNOWN

We need to handle DT_UNKNOWN earlier, not only during directory
recursion, so that the cached file_type value in the directory_entry
won't be used.

* src/c++17/fs_dir.cc (_Dir::advance(bool, error_code&)): Handle
d_type == DT_UNKNOWN immediately.
(_Dir::should_recurse(bool, error_code&)): Remove file_type::unknown
handling here.
* testsuite/27_io/filesystem/iterators/caching.cc: New test.

From-SVN: r270172

Make filesystem::path safe for self assignment

The standard says "If *this and p are the same object, has no effect."
Previously we ended up clearing the path.

* include/bits/fs_path.h (path::operator=(path&&)): Check for self
assignment.
* src/c++17/fs_path.cc (path::operator=(const path&)): Likewise.
* testsuite/27_io/filesystem/path/assign/copy.cc: Test self
assignment.

From-SVN: r270171

PR libstdc++/87431 re-adjust never-valueless optimizations

Avoid creating arbitrarily large objects on the stack when emplacing
trivially copyable objects into a variant. Currently we provide the
strong exception-safety guarantee for all trivially copyable types, by
constructing a second variant and then doing a non-throwing move
assignment from the temporary. This patch restricts that behaviour to
trivially copyable types that are no larger than 256 bytes. For larger
types the object will be emplaced directly into the variant, and if its
initialization throws then the variant becomes valueless.

Also implement Antony Polukhin's suggestion to whitelist specific types
that are not trivially copyable but can be efficiently move-assigned.
Emplacing those types will never cause a variant to become valueless.
The whitelisted types are: std::shared_ptr, std::weak_ptr,
std::unique_ptr, std::function, and std::any. Additionally,
std::basic_string, std::vector, and __gnu_debug::vector are whitelisted
if their allocator traits give them a non-throwing move assignment
operator. Specifically, this means std::string is whitelisted, but
std::pmr::string is not.

As part of this patch, additional if-constexpr branches are added for
the cases where the initialization is known to be non-throwing (so the
overhead of the try-catch block can be avoided) and where a scalar is
being produced by a potentially-throwing conversion operator (so that
the overhead of constructing and move-assigning a variant is avoided).
These changes should have no semantic effect, just better codegen.

PR libstdc++/87431 (again)
* include/bits/basic_string.h (__variant::_Never_valueless_alt):
Define partial specialization for basic_string.
* include/bits/shared_ptr.h (_Never_valueless_alt): Likewise for
shared_ptr and weak_ptr.
* include/bits/std_function.h (_Never_valueless_alt): Likewise for
function.
* include/bits/stl_vector.h (_Never_valueless_alt): Likewise for
vector.
* include/bits/unique_ptr.h (_Never_valueless_alt): Likewise for
unique_ptr.
* include/debug/vector (_Never_valueless_alt): Likewise for debug
vector.
* include/std/any (_Never_valueless_alt): Define explicit
specialization for any.
* include/std/variant (_Never_valueless_alt): Define primary template.
(__never_valueless): Use _Never_valueless_alt instead of
is_trivially_copyable.
(variant::emplace<N>(Args&&...)): Add special case for non-throwing
initializations to avoid try-catch overhead. Add special case for
scalars produced by potentially-throwing conversions. Use
_Never_valueless_alt instead of is_trivially_copyable for the
remaining strong exception-safety cases.
(variant::emplace<N>(initializer_list<U>, Args&&...)): Likewise.
* testsuite/20_util/variant/87431.cc: Run both test functions.
* testsuite/20_util/variant/exception_safety.cc: New test.
* testsuite/20_util/variant/run.cc: Use pmr::string instead of string,
so the variant becomes valueless.

From-SVN: r270170

Guard notes for -Waddress-of-packed-member on warning emission (PR c/89985)

gcc/c-family/ChangeLog:
PR c/89985
* c-warn.c (check_address_or_pointer_of_packed_member): Add
auto_diagnostic_group.  Guard inform calls by result of
warning_at call.

gcc/testsuite/ChangeLog:
PR c/89985
* c-c++-common/pr89985.c: New test.

From-SVN: r270169

[testsuite] PR71598: Fix testcases again

2019-04-05  Christophe Lyon  <christophe.lyon@linaro.org>

PR c/71598
* gcc.dg/torture/pr71598-1.c: dg-prune arm linker messages about
size of enums.
* gcc.dg/torture/pr71598-2.c: Likewise.

From-SVN: r270168

sched-deps.c (sched_macro_fuse_insns): Check return value of targetm.fixed_condition_code_regs.

        * sched-deps.c (sched_macro_fuse_insns): Check return value of
        targetm.fixed_condition_code_regs.

From-SVN: r270167

PR c++/89973 - -Waddress-of-packed-member ICE with invalid conversion.

PR c++/89973 - -Waddress-of-packed-member ICE with invalid conversion.
* c-warn.c (check_address_or_pointer_of_packed_member): Check the type
of RHS.

* g++.dg/warn/Waddress-of-packed-member2.C: New test.

From-SVN: r270166

re PR debug/89892 (gcc generates wrong debug information at -O2)

2019-04-05  Richard Biener  <rguenther@suse.de>

PR debug/89892
PR debug/89905
* tree-cfgcleanup.c (remove_forwarder_block): Always move
debug bind stmts but reset them if they are not valid at the
destination.

* gcc.dg/guality/pr89892.c: New testcase.
* gcc.dg/guality/pr89905.c: Likewise.
* gcc.dg/guality/loop-1.c: Likewise.

From-SVN: r270165

Fix ChangeLog entries.

contrib/ChangeLog:

2019-04-03  Martin Liska  <mliska@suse.cz>

PR translation/89936
* check-internal-format-escaping.py: Properly detect wrong
apostrophes.

gcc/ChangeLog:

2019-04-03  Martin Liska  <mliska@suse.cz>

PR translation/89936
* collect-utils.c (collect_execute): Use %< and %>, or %qs in
order to wrap keywords or arguments.
* collect2.c (main): Likewise.
(scan_prog_file): Likewise.
(scan_libraries): Likewise.
* common/config/riscv/riscv-common.c (riscv_subset_list::parsing_subset_version): Likewise.
(riscv_subset_list::parse_std_ext): Likewise.
* config/aarch64/aarch64.c (aarch64_override_options_internal): Likewise.
* config/arm/arm.c (arm_option_override): Likewise.
* config/cris/cris.c (cris_print_operand): Likewise.
* config/darwin-c.c (darwin_pragma_options): Likewise.
(darwin_pragma_unused): Likewise.
(darwin_pragma_ms_struct): Likewise.
* config/ft32/ft32.c (ft32_print_operand): Likewise.
* config/i386/i386.c (print_reg): Likewise.
(ix86_print_operand): Likewise.
* config/i386/xm-djgpp.h: Likewise.
* config/iq2000/iq2000.c (iq2000_print_operand): Likewise.
* config/m32c/m32c.c (m32c_option_override): Likewise.
* config/msp430/msp430.c (msp430_option_override): Likewise.
* config/nds32/nds32.c (nds32_option_override): Likewise.
* config/nvptx/mkoffload.c (main): Likewise.
* config/rx/rx.c (rx_print_operand): Likewise.
(valid_psw_flag): Likewise.
* config/vms/vms-c.c (vms_pragma_member_alignment): Likewise.
(vms_pragma_nomember_alignment): Likewise.
(vms_pragma_extern_model): Likewise.
* lto-wrapper.c (compile_offload_image): Likewise.
* omp-offload.c (oacc_parse_default_dims): Likewise.
* symtab.c (symtab_node::verify_base): Likewise.
* tlink.c (recompile_files): Likewise.
(start_tweaking): Likewise.
* tree-profile.c (parse_profile_filter): Likewise.

gcc/objc/ChangeLog:

2019-04-03  Martin Liska  <mliska@suse.cz>

PR translation/89936
* objc-act.c (objc_add_property_declaration): Use %< and %>, or %qs in
order to wrap keywords or arguments.
(objc_add_synthesize_declaration_for_property): Likewise.

From-SVN: r270164

Remove usage of apostrophes in error and warning messages (PR translation/89935).

2019-04-05  Martin Liska  <mliska@suse.cz>

PR translation/89935
* check-internal-format-escaping.py: Properly detect wrong
apostrophes.
2019-04-05  Martin Liska  <mliska@suse.cz>

PR translation/89935
* collect-utils.c (collect_execute): Use %< and %>, or %qs in
order to wrap keywords or arguments.
* collect2.c (main): Likewise.
(scan_prog_file): Likewise.
(scan_libraries): Likewise.
* common/config/riscv/riscv-common.c (riscv_subset_list::parsing_subset_version): Likewise.
(riscv_subset_list::parse_std_ext): Likewise.
* config/aarch64/aarch64.c (aarch64_override_options_internal): Likewise.
* config/arm/arm.c (arm_option_override): Likewise.
* config/cris/cris.c (cris_print_operand): Likewise.
* config/darwin-c.c (darwin_pragma_options): Likewise.
(darwin_pragma_unused): Likewise.
(darwin_pragma_ms_struct): Likewise.
* config/ft32/ft32.c (ft32_print_operand): Likewise.
* config/i386/i386.c (print_reg): Likewise.
(ix86_print_operand): Likewise.
* config/i386/xm-djgpp.h: Likewise.
* config/iq2000/iq2000.c (iq2000_print_operand): Likewise.
* config/m32c/m32c.c (m32c_option_override): Likewise.
* config/msp430/msp430.c (msp430_option_override): Likewise.
* config/nds32/nds32.c (nds32_option_override): Likewise.
* config/nvptx/mkoffload.c (main): Likewise.
* config/rx/rx.c (rx_print_operand): Likewise.
(valid_psw_flag): Likewise.
* config/vms/vms-c.c (vms_pragma_member_alignment): Likewise.
(vms_pragma_nomember_alignment): Likewise.
(vms_pragma_extern_model): Likewise.
* lto-wrapper.c (compile_offload_image): Likewise.
* omp-offload.c (oacc_parse_default_dims): Likewise.
* symtab.c (symtab_node::verify_base): Likewise.
* tlink.c (recompile_files): Likewise.
(start_tweaking): Likewise.
* tree-profile.c (parse_profile_filter): Likewise.
2019-04-05  Martin Liska  <mliska@suse.cz>

PR translation/89935
* objc-act.c (objc_add_property_declaration): Use %< and %>, or %qs in
order to wrap keywords or arguments.
(objc_add_synthesize_declaration_for_property): Likewise.

From-SVN: r270163

Make FMA code cope with redundant negates (PR89956)

This patch fixes a case in which, due to forced missed optimisations
in earlier passes, we have:

    _1 = a * b
    _2 = -_1
    _3 = -_1
    _4 = _2 + _3

and treated _4 as two FNMA candidates, once via _2 and once via _3.

2019-04-05  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR tree-optimization/89956
* tree-ssa-math-opts.c (convert_mult_to_fma): Protect against
multiple negates of the same value.

gcc/testsuite/
PR tree-optimization/89956
* gfortran.dg/pr89956.f90: New test.

From-SVN: r270162

PR c++/89948 - ICE with break in statement-expr.

* constexpr.c (cxx_eval_statement_list): Jumping out of a
statement-expr is non-constant.

From-SVN: r270161

PR c++/89966 - error with non-type auto tparm.

My patch for PR 86932 broke this testcase by passing tf_partial to
coerce_template_template_parms, which prevented do_auto_deduction from
actually replacing the auto.

* pt.c (do_auto_deduction): Clear tf_partial.

From-SVN: r270160

PR c++/86986 - ICE with TTP with parameter pack.

Three separate issues were breaking this testcase.  One, we were trying to
look at the type of a template template parameter to see if it's a valid
non-type template parameter.  Two, we were treating a parameter pack named
in the type of a template parameter pack of a TTP pack as being one of the
packs expanded by the outer pack.  Three, we weren't supplying all the
necessary levels of template arguments when TTP matching.

* pt.c (coerce_template_parameter_pack): Only look at the type of a
non-type parameter pack.
(fixed_parameter_pack_p_1): Don't recurse into the type of a
non-type parameter pack.
(coerce_template_template_parms): Call add_outermost_template_args.

From-SVN: r270159