aarch64.c (aarch_macro_fusion_pair_p): Don't access prev before checking it for NULLness in the...

2018-09-24  Andrew Pinski  <apinski@marvell.com>

        *  config/aarch64/aarch64.c (aarch_macro_fusion_pair_p): Don't
        access prev before checking it for NULLness in the
        AARCH64_FUSE_CMP_BRANCH case.

From-SVN: r264548

codecvt.xml: Move link to "UTF-8 and Unicode FAQ" to https.

* doc/xml/manual/codecvt.xml: Move link to "UTF-8 and Unicode FAQ"
to https.

From-SVN: r264547

libgo: update to Go 1.11

    Reviewed-on: https://go-review.googlesource.com/136435

gotools/:
* Makefile.am (mostlyclean-local): Run chmod on check-go-dir to
make sure it is writable.
(check-go-tools): Likewise.
(check-vet): Copy internal/objabi to check-vet-dir.
* Makefile.in: Rebuild.

From-SVN: r264546

cp-tree.h (build_noexcept_spec, [...]): Adjust declarations.

* cp-tree.h (build_noexcept_spec, add_exception_specifier): Adjust
declarations.
* except.c (build_noexcept_spec): Change the type of the complain
parameter to tsubst_flags_t.
* typeck2.c (add_exception_specifier): Likewise.

From-SVN: r264543

re PR c++/85070 (ICE on C++ code: in lazily_declare_fn, at cp/method.c:2409)

/cp
2018-09-24  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/85070
* method.c (lazily_declare_fn): During error-recovery add_method
may return false.

/testsuite
2018-09-24  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/85070
* g++.dg/cpp0x/pr85070.C: New.

From-SVN: r264541

i386: Insert ENDBR before the profiling counter call

ENDBR must be the first instruction of a function.  This patch queues
ENDBR if we need to put the profiling counter call before the prologue
and generate ENDBR before the profiling counter call.

gcc/

PR target/82699
* config/i386/i386.c (rest_of_insert_endbranch): Set
endbr_queued_at_entrance to true and don't insert ENDBR if
x86_function_profiler will be called.
(x86_function_profiler): Insert ENDBR if endbr_queued_at_entrance
is true.
* config/i386/i386.h (machine_function): Add
endbr_queued_at_entrance.

gcc/testsuite/

PR target/82699
* gcc.target/i386/pr82699-1.c: New file.
* gcc.target/i386/pr82699-2.c: Likewise.
* gcc.target/i386/pr82699-3.c: Likewise.
* gcc.target/i386/pr82699-4.c: Likewise.
* gcc.target/i386/pr82699-5.c: Likewise.
* gcc.target/i386/pr82699-6.c: Likewise.

From-SVN: r264540

re PR fortran/87397 (Clobbering intent(out) variables caused regression in OpenCoarrays testsuite)

2018-09-24  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87397
* gfc_conv_procedure_call: Do not add clobber on INTENT(OUT)
for variables in an associate statement.

2018-09-24  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87401
* gfortran.dg/intent_out_12.f90: New test.

From-SVN: r264539

re PR target/86952 (Avoid jump table for switch statement with -mindirect-branch=thunk)

[testsuite]

2018-09-24  Will Schmidt  <will_schmidt@vnet.ibm.com>

PR testsuite/86952
* gcc.target/powerpc/p8-vec-xl-xst-v2.c: Add and
update expected codegen

From-SVN: r264538

Change EQ_ATTR_ALT to support up to 64 alternatives

On S/390 there is a need to support more than 32 instruction
alternatives per define_insn.  Currently this is not explicitly
prohibited or unsupported: MAX_RECOG_ALTERNATIVES is equal 35, and,
futhermore, the related code uses uint64_t for bitmaps in most places.

However, genattrtab contains the logic to convert (eq_attr "attribute"
"value") RTXs to (eq_attr_alt bitmap) RTXs, where bitmap contains
alternatives, whose "attribute" has the corresponding "value".
Unfortunately, bitmap is only 32 bits.

When adding the 33rd alternative, this led to (eq_attr "type" "larl")
becoming (eq_attr_alt -1050625 1), where -1050625 == 0xffeff7ff.  The
cleared bits 12, 21 and 32 correspond to two existing and one newly
added insn of type "larl".  compute_alternative_mask sign extended this
to 0xffffffffffeff7ff, which contained non-existent alternatives, and
this made simplify_test_exp fail with "invalid alternative specified".

I'm not sure why it didn't fail the same way before, since the top bit,
which led to sign extension, should have been set even with 32
alternatives.  Maybe simplify_test_exp was not called for "type"
attribute for some reason?

This patch widens EQ_ATTR_ALT bitmap to 64 bits, making it possible to
gracefully handle up to 64 alternatives.  It eliminates the problem with
the 33rd alternative on S/390.

gcc/ChangeLog:

2018-09-24  Ilya Leoshkevich  <iii@linux.ibm.com>

* genattrtab.c (mk_attr_alt): Use alternative_mask.
(attr_rtx_1): Adjust caching to match the new EQ_ATTR_ALT field
        types.
(check_attr_test): Use alternative_mask.
(get_attr_value): Likewise.
(compute_alternative_mask): Use alternative_mask and XWINT.
(make_alternative_compare): Use alternative_mask.
(attr_alt_subset_p): Use XWINT.
(attr_alt_subset_of_compl_p): Likewise.
(attr_alt_intersection): Use alternative_mask and XWINT.
(attr_alt_union): Likewise.
(attr_alt_complement): Use HOST_WIDE_INT and XWINT.
        (mk_attr_alt): Use alternative_mask and HOST_WIDE_INT.
(simplify_test_exp): Use alternative_mask and XWINT.
(write_test_expr): Use alternative_mask and XWINT, adjust bit
        number calculation to support 64 bits.  Generate code that
        checks 64-bit masks.
(main): Use alternative_mask.
* rtl.def (EQ_ATTR_ALT): Change field types from ii to ww.

From-SVN: r264537

[OpenACC] update fortran nested parallelism error messages

gcc/fortran/
* openmp.c (resolve_oacc_loop_blocks):

gcc/testsuite/
* gfortran.dg/goacc/nested-parallelism.f90: New test.

From-SVN: r264536

S/390: Fix conditional returns on z196+

S/390 epilogue ends with (parallel [(return) (use %r14)]) instead of
the more usual (return) or (simple_return).  This sequence is not
recognized by the conditional return logic in try_optimize_cfg ().

This was introduced for processors older than z196, where it is
sometimes profitable to use call-clobbered register for returning
instead of %r14.  On newer processors we always return via %r14,
for which the fact that it's used is already reflected by
EPILOGUE_USES.  In this case a simple (return) suffices.

This patch changes return_use () to emit simple (return)s when
returning via %r14.  The resulting sequences are recognized by the
conditional return logic in try_optimize_cfg ().

gcc/ChangeLog:

2018-09-24  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/80080
* config/s390/s390.c (s390_emit_epilogue): Do not use PARALLEL
RETURN+USE when returning via %r14.

gcc/testsuite/ChangeLog:

2018-09-24  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/80080
* gcc.target/s390/risbg-ll-3.c: Expect conditional returns.
* gcc.target/s390/zvector/vec-cmp-2.c: Likewise.

From-SVN: r264535

Improve colorization legend in gcov reports.

2018-09-24  Martin Liska  <mliska@suse.cz>

* gcov.c (output_lines): Print colorization legend
for both flag_use_colors and flag_use_hotness_colors.
Reword the help.

From-SVN: r264531

Improve location information of -Wcoverage-mismatch.

2018-09-24  Martin Liska  <mliska@suse.cz>

* coverage.c (get_coverage_counts): Use warning_at
with current_function_decl location. Use %qD in warning
message.

From-SVN: r264530

Guard memory block allocation.

2018-09-24  Martin Liska  <mliska@suse.cz>

* memory-block.h (memory_block_pool::release): Annotate with
valgrind that the memory is not accessible.

From-SVN: r264529

Unpoison variable partition properly (PR sanitizer/85774).

2018-09-24  Martin Liska  <mliska@suse.cz>

PR sanitizer/85774
* asan.c: Make asan_handled_variables extern.
* asan.h: Likewise.
* cfgexpand.c (expand_stack_vars): Make sure
a representative is unpoison if another
variable in the partition is handled by
use-after-scope sanitization.
2018-09-24  Martin Liska  <mliska@suse.cz>

PR sanitizer/85774
* g++.dg/asan/pr85774.C: New test.

From-SVN: r264528

policy_data_structures_biblio.xml: Update link to "Priority Queues and the STL".

* doc/xml/manual/policy_data_structures_biblio.xml: Update link
to "Priority Queues and the STL".

From-SVN: r264527

[PR87054] adjust testcase for 32-bit x86

The test assumed __int128 to be available whenever __SSE__ was
defined, but this assumption doesn't hold on 32-bit x86.  Fixed.

for  gcc/testsuite/ChangeLog

PR middle-end/87054
* gcc.dg/pr87054.c: Adjust for no __int128 on x86.

From-SVN: r264526

re PR ada/87396 (ada/gcc-interface/decl.c:8798:53:Value Conversion Issue: implicit conversion from 'int' to 'char' changes value from 132 to -124: -Wconstant-conversion)

PR ada/87396
* fe.h (Get_Attribute_Definition_Clause): Use 'unsigned char' instead
of 'char' as the type of the second parameter.

From-SVN: r264525

re PR middle-end/63155 (memory hog)

2018-09-24  Richard Biener  <rguenther@suse.de>

PR tree-optimization/63155
* tree-ssa-propagate.c (add_ssa_edge): Avoid adding PHIs to
the worklist when the edge of the respective argument isn't
executable.

From-SVN: r264523

Daily bump.

From-SVN: r264522

re PR fortran/87397 (Clobbering intent(out) variables caused regression in OpenCoarrays testsuite)

2018-09-23  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87397
* gfc_conv_procedure_call: Do not add clobber on INTENT(OUT)
for variables having the dimension attribute.

2018-09-23  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87395
* gfortran.dg/intent_out_11.f90: New test.

From-SVN: r264518

i386.h (enum reg_class): Rename MASK_REGS to ALL_MASK_REGS and MASK_EVEX_REGS to MASK_REGS.

* config/i386/i386.h (enum reg_class): Rename MASK_REGS to
ALL_MASK_REGS and MASK_EVEX_REGS to MASK_REGS.
(MASK_CLASS_P): Update for rename.
(MAYBE_MASK_CLASS_P): Ditto.
(REG_CLASS_NAMES): Update.
(REG_CLASS_CONTENT): Update.
* config/i386/i386.c (regclass_map): Update for MASK_REG
and ALL_MASK_REGS rename.
* config/i386/constraints.md (Yk): Update for rename.
(k): Ditto.

* config/i386/i386.h (enum reg_class): Remove
EVEX_SSE_REGS and MOD4_SSE_REGS.
(REG_CLASS_NAMES): Update.
(REG_CLASS_CONTENT): Update.
* config/i386/i386.c (regclass_map): Declare AVX-512 SSE
registers as ALL_SSE_REGS.
(ix86_additional_allocno_class_p): Remove.
(TARGET_ADDITIONAL_ALLOCNO_CLASS_P): Remove.
(ix86_register_priority): Lower priority of EVEX SSE registers.
Use IN_RANGE macro where appropriate.
(ix86_hard_regno_mode_ok): Merge AVX-5124FMAPS and
AVX-5124VNNIW checks.
(ix86_modes_tieable_p): Tie 512-bit SSE modes.
* config/i386/sse.md (avx5124fmaddps_4fmaddps)
(avx5124fmaddps_4fmaddps_mask, avx5124fmaddps_4fmaddps_maskz)
(avx5124fmaddps_4fmaddss, avx5124fmaddps_4fmaddss_mask)
(avx5124fmaddps_4fmaddss_maskz, avx5124fmaddps_4fnmaddps)
(avx5124fmaddps_4fnmaddps_mask, avx5124fmaddps_4fnmaddps_maskz)
(avx5124fmaddps_4fnmaddss, avx5124fmaddps_4fnmaddss_mask)
(avx5124fmaddps_4fnmaddss_maskz, avx5124vnniw_vp4dpwssd)
(avx5124vnniw_vp4dpwssd_mask, avx5124vnniw_vp4dpwssd_maskz)
(avx5124vnniw_vp4dpwssds, avx5124vnniw_vp4dpwssds_mask)
(avx5124vnniw_vp4dpwssds_maskz): Use "v" instead of "Yh" constraint.
* config/i386/constraints.md (Yh): Remove.

* config/i386/i386.c (regclass_map): Declare integer REX registers
        as GENERAL_REGS.

From-SVN: r264516

using_exceptions.xml: Move boost.orgs link to https.

* doc/xml/manual/using_exceptions.xml: Move boost.orgs link to
https.

From-SVN: r264515

* doc/service.texi (Service): Switch the fsf.org link to https.

From-SVN: r264514

Do array index calculations in gfc_array_index_type

It was recently noticed that for a few of the coarray intrinsics array
index calculations were done in integer_type_node instead of
gfc_array_index_type.  This patch fixes this.

Regtested on x86_64-pc-linux-gnu.

gcc/fortran/ChangeLog:

2018-09-23  Janne Blomqvist  <jb@gcc.gnu.org>

* trans-expr.c (gfc_caf_get_image_index): Do array index
calculations in gfc_array_index_type.
* trans-intrinsic.c (conv_intrinsic_event_query): Likewise.
* trans-stmt.c (gfc_trans_lock_unlock): Likewise.
(gfc_trans_event_post_wait): Likewise.

gcc/testsuite/ChangeLog:

2018-09-23  Janne Blomqvist  <jb@gcc.gnu.org>

* gfortran.dg/coarray_lib_alloc_4.f90: Fix scan patterns.
* gfortran.dg/coarray_lock_7.f90: Likewise.

From-SVN: r264513

re PR fortran/87395 (ICE in in lookup_field_for_decl with whizard)

2018-09-23  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87395
* gfc_conv_procedure_call: Reformat comments slightly. Do not add
clobber on INTENT(OUT) for saved variables.

2018-09-23  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87395
* gfortran.dg/intent_out_10.f90: New test.

From-SVN: r264512

Daily bump.

From-SVN: r264510

re PR middle-end/41453 (use INTENT(out) for optimization)

2018-09-22  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/41453
* trans.h (gfc_conv_expr_reference): Add optional argument
add_clobber to prototype.
(gfc_conv_procedure_call):  Set add_clobber argument to
gfc_conv_procedure_reference to true for scalar, INTENT(OUT),
non-pointer, non-allocatable, non-dummy variables whose type
is neither BT_CHARACTER, BT_DERIVED or BT_CLASS, but only if
the procedure is not elemental.
* trans-expr.c (gfc_conv_procedure_reference): Add clobber
statement before call if add_clobber is set.

2018-09-22  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/41453
* gfortran.dg/intent_optimize_2.f90: New test.

From-SVN: r264506

re PR fortran/87318 (gfortran.dg/dtio_1.f90 is invalid)

2018-09-22  Jerry DeLisle  <jvdelisle@gcc.gnu.org>

PR fortran/87318
* gfortran.dg/dtio_1.f90: Update test to valid code.

From-SVN: r264505

re PR target/86798 (nds32 port needs updating for CVE-2017-5753)

PR target/86798
* config/nds32/nds32.c (TARGET_HAVE_SPECULATION_SAFE_VALUE):
Define to speculation_safe_value_not_needed.

From-SVN: r264503

re PR fortran/85603 (ICE with character array substring assignment)

2018-09-22  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/85603
* trans-array.c (gfc_alloc_allocatable_for_assignment): Test
the charlen backend_decl before using the VAR_P macro.

2018-09-22  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/85603
* gfortran.dg/deferred_character_23.f90 : New test.

From-SVN: r264502

Daily bump.

From-SVN: r264500

re PR libstdc++/87135 ([C++17] unordered containers violate iterator validity requirements)

2018-09-21  François Dumont  <fdumont@gcc.gnu.org>

PR libstdc++/87135
* src/c++11/hashtable_c++0x.cc (_Prime_rehash_policy::_M_next_bkt):
Use __builtin_floor to compute _M_next_resize.
* testsuite/23_containers/unordered_set/hash_policy/71181.cc: Adapt.
* testsuite/23_containers/unordered_set/hash_policy/prime_rehash.cc:
Adapt.

From-SVN: r264494

gimple-ssa-evrp.c (evrp_dom_walker::cleanup): Call vr_values::cleanup_edges_and_switches.

* gimple-ssa-evrp.c (evrp_dom_walker::cleanup): Call
vr_values::cleanup_edges_and_switches.
* tree-vrp.c (to_remove_edges, to_update_switch_stmts): Moved into
vr_values class.
(identify_jump_threads): Remove EDGE_IGNORE handling.
(execute_vrp): Move handling of to_remove_edges and
to_update_switch_stmts into vr_values class member functions.
* tree-vrp.h (switch_update, to_remove_edges): Remove declarations.
(to_update_switch_stmts): Likewise.
* vr-values.c: Include cfghooks.h.
(vr_values::vr_values): Initialize to_remove_edges and
to_update_switch_stmts.
(vr_values::~vr_values): Verify to_remove_edges and
to_update_switch_stmts are empty.
(vr_values::simplify_switch_using_ranges): Set EDGE_IGNORE as needed.
(vr_values::cleanup_edges_and_switches): New member function.
* vr-values.h (vr_values): Add cleanup_edges_and_switches member
function.  Add new data members.

* gcc.dg/tree-ssa/vrp113.c: Disable EVRP.
* gcc.dg/tree-ssa/vrp120.c: New test.

From-SVN: r264491

Document that attribute noreturn inhibits tail call optimization

PR middle-end/81035
* doc/extend.texi (Common Function Attributes): Mention that
noreturn suppresses tail call optimization.

From-SVN: r264490

PR c++/87372 - __func__ constexpr evaluation.

* constexpr.c (maybe_constant_init_1): Pass false for strict down to
cxx_eval_outermost_constant_expr.

* g++.dg/cpp1y/func_constexpr2.C: New test.

From-SVN: r264489

Use vectored writes when reporting errors and warnings.

When producing error and warning messages, libgfortran writes a
message by using many system calls.  By using vectored writes (the
POSIX writev function) when available and feasible to use without
major surgery, we reduce the chance that output gets intermingled with
other output to stderr.

In practice, this is done by introducing a new function estr_writev in
addition to the existing estr_write.  In order to use this, the old
st_vprintf is removed, replaced by direct calls of vsnprintf, allowing
more message batching.

Regtested on x86_64-pc-linux-gnu.

libgfortran/ChangeLog:

2018-09-21  Janne Blomqvist  <jb@gcc.gnu.org>

* config.h.in: Regenerated.
* configure: Regenerated.
* configure.ac: Check for writev and sys/uio.h.
* libgfortran.h: Include sys/uio.h.
(st_vprintf): Remove prototype.
(struct iovec): Define if not available.
(estr_writev): New prototype.
* runtime/backtrace.c (error_callback): Use estr_writev.
* runtime/error.c (ST_VPRINTF_SIZE): Remove.
(estr_writev): New function.
(st_vprintf): Remove.
(gf_vsnprintf): New function.
(ST_ERRBUF_SIZE): New macro.
(st_printf): Use vsnprintf.
(os_error): Use estr_writev.
(runtime_error): Use vsnprintf and estr_writev.
(runtime_error_at): Likewise.
(runtime_warning_at): Likewise.
(internal_error): Use estr_writev.
(generate_error_common): Likewise.
(generate_warning): Likewise.
(notify_std): Likewise.
* runtime/pause.c (pause_string): Likewise.
* runtime/stop.c (report_exception): Likewise.
(stop_string): Likewise.
(error_stop_string): Likewise.

From-SVN: r264487

re PR fortran/77325 (ICE in gimplify_var_or_parm_decl, at gimplify.c:1933)

2018-09-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/77325
* trans-array.c (gfc_alloc_allocatable_for_assignment): If the
rhs has a charlen expression, convert that and use it.
* trans-expr.c (gfc_trans_assignment_1): The rse.pre for the
assignment of deferred character array vars to a realocatable
lhs should not be added to the exterior block since vector
indices, for example, generate temporaries indexed within the
loop.

2018-09-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/77325
* gfortran.dg/deferred_character_22.f90 : New test.

From-SVN: r264486

re PR fortran/87359 (pointer being freed was not allocated)

2018-09-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87359
* trans-stmt.c (gfc_trans_allocate): Don't deallocate alloc
components if must_finalize is set for expr3.

2018-09-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87359
* gfortran.dg/finalize_33.f90 : New test.

From-SVN: r264485

* constexpr.c (cxx_eval_outermost_constant_expr): Update comment.

From-SVN: r264484

Un-split hypot<long double> tests

Remove the hypot-long-double.cc file that used dg-xfail-run-if and
simply use the lower tolerance for double if long double is not larger
than double.

* testsuite/26_numerics/headers/cmath/hypot-long-double.cc: Remove.
* testsuite/26_numerics/headers/cmath/hypot.cc: Restore test for
long double unconditionally, but use lower tolerance when
sizeof(long double) == sizeof(double).

From-SVN: r264483

dumpfile.c: fix stray dump_loc output (PR tree-optimization/87309)

In r262891 I reimplemented this call:
  dump_printf_loc (MSG_NOTE, loc, "=== %s ===\n", name);
in dump_begin_scope to use direct calls to dump_loc:
  if (dump_file)
    {
      dump_loc (MSG_NOTE, dump_file, loc.get_location_t ());
      fprintf (dump_file, "=== %s ===\n", name);
    }

  if (alt_dump_file)
   {
     dump_loc (MSG_NOTE, alt_dump_file, loc.get_location_t ());
     fprintf (alt_dump_file, "=== %s ===\n", name);
   }

However ::dump_loc doesn't filter with pflags and alt_flags.

This lead to stray output of the form:
  test.cpp:1:6: note: test.cpp:1:11: note:
when using -fopt-info with "optimized" or "missed".

This patch adds this missing filtering, eliminating the stray partial
note output.

gcc/ChangeLog:
PR tree-optimization/87309
* dumpfile.c (dump_context::begin_scope): Filter the dump_loc
calls with pflags and alt_flags.
(selftest::test_capture_of_dump_calls): Add test of interaction of
MSG_OPTIMIZED_LOCATIONS with AUTO_DUMP_SCOPE.

gcc/testsuite/ChangeLog:
PR tree-optimization/87309
* gcc.dg/pr87309.c: New test.

From-SVN: r264481

config.gcc: factorize and comment inclusion of vxworks-dummy.h

2018-09-21  Olivier Hainque  <hainque@adacore.com>

* gcc/config.gcc: Factorize and comment inclusion of
vxworks-dummy.h.

From-SVN: r264480

Leverage cacheTextUpdate for __clear_cache on VxWorks

2018-09-21  Alexandre Oliva  <oliva@adacore.com>

libgcc/
* config/vxcache.c: New file.  Provide __clear_cache, based on
the cacheTextUpdate VxWorks service.
* config/t-vxworks (LIB2ADD): Add vxcache.c.
(LIB2FUNCS_EXCLUDE): Add _clear_cache.
* config/t-vxwoks7: Likewise.
gcc/
* config/vxworks.h (CLEAR_INSN_CACHE): #define to 1.

From-SVN: r264479

Cleanup handling of libgcc and libc_internal for VxWorks

2018-09-21  Olivier Hainque  <hainque@adacore.com>

* config/vxworks.h (VXWORKS_LIBGCC_SPEC): Remove -lc_internal.
Merge block comment with the one ahead of VXWORKS_LIBS_RTP. Then:
(VXWORKS_LIBS_RTP): Minor reordering.

From-SVN: r264477

Leverage STARTFILE_PREFIX_SPEC for VxWorks7

2018-09-21  Olivier Hainque  <hainque@adacore.com>

* config/vxworks.h (STARTFILE_PREFIX_SPEC): Define.
(VXWORKS_LIBS_DIR_RTP): Remove definition and use.

From-SVN: r264476

Account for TARGET_VXWORKS64 in default SIZE/PTRDIFF_TYPE for vxworks

2018-09-21  Olivier Hainque  <hainque@adacore.com>

* config/vxworks.h (SIZE_TYPE): Account for TARGET_VXWORKS64.
(PTRDIFF_TYPE): Likewise.

From-SVN: r264475

Introduce TARGET_VXWORKS64 for VxWorks 64bit ports

2018-09-21  Olivier Hainque  <hainque@adacore.com>

* config.gcc: Enforce def of TARGET_VXWORKS64 to 1 from
triplet, similar to support for VxWorks7.
* config/vxworks-dummy.h: Provide a default definition
of TARGET_VXWORKS64 to 0.

From-SVN: r264474

Fix whitespace vs tab in entry

From-SVN: r264473

Move default #define TARGET_VXWORKS7 to vxworks-dummy.h

2018-09-21  Olivier Hainque  <hainque@adacore.com>

* config/vxworks.h (TARGET_VXWORKS7): Move default definition ...
* config/vxworks-dummy.h: here.

From-SVN: r264472

Prepend vxworks-dummy.h to tm_file for powerpc*

2018-09-20  Olivier Hainque  <hainque@adacore.com>

* config.gcc (powerpc*-*-*): Prepend vxworks-dummy.h to tm_file.

From-SVN: r264471

[aarch64] Add HiSilicon tsv110 CPU support

This patch adds HiSilicon's an mcpu: tsv110, which supports v8_4A.
It has been tested on aarch64 and no regressions from this patch.

2018-09-21  Shaokun Zhang  <zhangshaokun@hisilicon.com>
            Bo Zhou  <zbo.zhou@hisilicon.com>

* config/aarch64/aarch64-cores.def (tsv110): New CPU.
* config/aarch64/aarch64-tune.md: Regenerated.
* doc/invoke.texi (AArch64 Options/-mtune): Add "tsv110".
* config/aarch64/aarch64.c (tsv110_tunings): New tuning table.
* config/aarch64/aarch64-cost-tables.h: Add "tsv110" extra costs.

Co-Authored-By: Bo Zhou <zbo.zhou@hisilicon.com>
From-SVN: r264470

Fix co-array allocation

The Fortran front-end has a bug in which it uses "int" values for "size_t"
parameters.  I don't know why this isn't problem for all 64-bit architectures,
but GCN ends up with the data in the wrong argument register and/or stack slot,
and bad things happen.

This patch corrects the issue by setting the correct type.

2018-09-21  Andrew Stubbs  <ams@codesourcery.com>
            Kwok Cheung Yeung  <kcy@codesourcery.com>

gcc/fortran/
* trans-expr.c (gfc_trans_structure_assign): Ensure that the first
argument of a call to _gfortran_caf_register is of size_type_node.
* trans-intrinsic.c (conv_intrinsic_event_query): Convert computed
index to a size_type_node type.
* trans-stmt.c (gfc_trans_event_post_wait): Likewise.

Co-Authored-By: Kwok Cheung Yeung <kcy@codesourcery.com>
From-SVN: r264468

Propagate address spaces to builtins.

At present, pointers passed to builtin functions, including atomic operators,
are stripped of their address space properties.  This doesn't seem to be
deliberate, it just omits to copy them.

Not only that, but it forces pointer sizes to Pmode, which isn't appropriate
for all address spaces.

This patch attempts to correct both issues.  It works for GCN atomics and
GCN OpenACC gang-private variables.

2018-09-21  Andrew Stubbs  <ams@codesourcery.com>
    Julian Brown  <julian@codesourcery.com>

gcc/
* builtins.c (get_builtin_sync_mem): Handle address spaces.

Co-Authored-By: Julian Brown <julian@codesourcery.com>
From-SVN: r264467

c-ada-spec.c: Include diagnostic.h.

* c-ada-spec.c: Include diagnostic.h.
(dump_ada_declaration) <RECORD_TYPE>: Issue a warning on packed layout.

From-SVN: r264466

rs6000.c (rs6000_function_ok_for_sibcall): Return false if the call takes a static chain.

* config/rs6000/rs6000.c (rs6000_function_ok_for_sibcall): Return false
if the call takes a static chain.

From-SVN: r264463

Remove arc profile histogram in non-LTO mode.

2018-09-21  Martin Liska  <mliska@suse.cz>

* auto-profile.c (autofdo_source_profile::read): Do not
set sum_all.
(read_profile): Do not add working sets.
(read_autofdo_file): Remove sum_all.
(afdo_callsite_hot_enough_for_early_inline): Remove const
qualifier.
* coverage.c (struct counts_entry): Remove gcov_summary.
(read_counts_file): Read new GCOV_TAG_OBJECT_SUMMARY,
do not support GCOV_TAG_PROGRAM_SUMMARY.
(get_coverage_counts): Remove summary and expected
arguments.
* coverage.h (get_coverage_counts): Likewise.
* doc/gcov-dump.texi: Remove -w option.
* gcov-dump.c (dump_working_sets): Remove.
(main): Do not support '-w' option.
(print_usage): Likewise.
(tag_summary): Likewise.
* gcov-io.c (gcov_write_summary): Do not dump
histogram.
(gcov_read_summary): Likewise.
(gcov_histo_index): Remove.
(gcov_histogram_merge): Likewise.
(compute_working_sets): Likewise.
* gcov-io.h (GCOV_TAG_OBJECT_SUMMARY): Mark
it not obsolete.
(GCOV_TAG_PROGRAM_SUMMARY): Mark it obsolete.
(GCOV_TAG_SUMMARY_LENGTH): Adjust.
(GCOV_HISTOGRAM_SIZE): Remove.
(GCOV_HISTOGRAM_BITVECTOR_SIZE): Likewise.
(struct gcov_summary): Simplify rapidly just
to runs and sum_max fields.
(gcov_histo_index): Remove.
(NUM_GCOV_WORKING_SETS): Likewise.
(compute_working_sets): Likewise.
* gcov-tool.c (print_overlap_usage_message): Remove
trailing empty line.
* gcov.c (read_count_file): Read GCOV_TAG_OBJECT_SUMMARY.
(output_lines): Remove program related line.
* ipa-profile.c (ipa_profile): Do not consider GCOV histogram.
* lto-cgraph.c (output_profile_summary): Do not stream GCOV
histogram.
(input_profile_summary): Do not read it.
(merge_profile_summaries): And do not merge it.
(input_symtab): Do not call removed function.
* modulo-sched.c (sms_schedule): Do not print sum_max.
* params.def (HOT_BB_COUNT_FRACTION): Reincarnate param that was
removed when histogram method was invented.
(HOT_BB_COUNT_WS_PERMILLE): Mention that it's used only in LTO
mode.
* postreload-gcse.c (eliminate_partially_redundant_load): Fix
GCOV coding style.
* predict.c (get_hot_bb_threshold): Use HOT_BB_COUNT_FRACTION
and dump selected value.
* profile.c (add_working_set): Remove.
(get_working_sets): Likewise.
(find_working_set): Likewise.
(get_exec_counts): Do not work with working sets.
(read_profile_edge_counts): Do not inform as sum_max is removed.
(compute_branch_probabilities): Likewise.
(compute_value_histograms): Remove argument for call of
get_coverage_counts.
* profile.h: Do not make gcov_summary const.
2018-09-21  Martin Liska  <mliska@suse.cz>

* libgcov-driver.c (crc32_unsigned): Remove.
(gcov_histogram_insert): Likewise.
(gcov_compute_histogram): Likewise.
(compute_summary): Simplify rapidly.
(merge_one_data): Do not handle PROGRAM_SUMMARY tag.
(merge_summary): Rapidly simplify.
(dump_one_gcov): Ignore gcov_summary.
(gcov_do_dump): Do not handle program summary, it's not
used.
* libgcov-util.c (tag_summary): Remove.
(read_gcda_finalize): Fix coding style.
(read_gcda_file): Initialize curr_object_summary.
(compute_summary): Remove.
(calculate_overlap): Remove settings of run_max.

From-SVN: r264462

[NDS32] Sync glibc and kernel structure, all use _rt_sigframe.

libgcc/
* config/nds32/linux-unwind.h (struct _rt_sigframe): Use struct
ucontext_t type instead.
(nds32_fallback_frame_state): Remove struct _sigframe statement.

From-SVN: r264461

[NDS32] Add t-nds32-glibc file.

libgcc/
* config/nds32/t-nds32-glibc: New file.

From-SVN: r264460

[NDS32] Turn off default option -mtlsdesc-trampoline.

gcc/
* config.gcc (nds32*-*-*): Set TARGET_DEFAULT_TLSDESC_TRAMPOLINE=0.

From-SVN: r264459

re PR tree-optimization/86990 (wrong code at -O2 on x86_64-linux-gnu in 64-bit mode)

PR tree-optimization/86990
* gimple-ssa-store-merging.c (imm_store_chain_info:coalesce_immediate):
Check that the entire merged store group is made of constants only for
overlapping stores.

From-SVN: r264458

Daily bump.

From-SVN: r264456

PR c++/87109 - wrong ctor with maybe-rvalue semantics.

* call.c (build_user_type_conversion_1): Refine the maybe-rvalue
check to only return if we're converting the return value to a base
class.

* g++.dg/cpp0x/ref-qual19.C: Adjust the expected results.
* g++.dg/cpp0x/ref-qual20.C: New test.

From-SVN: r264452

Fix and document -r option

The option has existed and been working for years,
make sure it implies the right extra options, and list
it in the documentation.

2018-09-20  Allan Sandfeld Jensen  <allan.jensen@qt.io>

gcc/
* gcc.c (LINK_COMMAND_SPEC): Handle -r like -nostdlib.
(VTABLE_VERIFICATION_SPEC): Likewise.
(SANITIZER_EARLY_SPEC): Likewise.
(SANITIZER_SPEC): Likewise.
* config/darwin.h (LINK_COMMAND_SPEC): Likewise.
* doc/invoke.texi (Link Options): Document -r.

gcc/cp/
* g++spec.c (lang_specific_driver): Handle -r like -nostdlib.

gcc/fortran/
* gfortranspec.c (lang_specific_driver): Handle -r like -nostdlib.

gcc/go/
* gospec.c (lang_specific_driver): Handle -r like -nostdlib.

gcc/testsuite/
* g++.dg/ipa/pr64059.C: Removed now redundant -nostdlib.
* g++.dg/lto/20081109-1_0.C: Likewise.
* g++.dg/lto/20090302_0.C: Likewise.
* g++.dg/lto/pr45621_0.C: Likewise.
* g++.dg/lto/pr60567_0.C: Likewise.
* g++.dg/lto/pr62026.C: Likewise.
* gcc.dg/lto/pr45736_0.c: Likewise.
* gcc.dg/lto/pr52634_0.c: Likewise.
* gfortran.dg/lto/20091016-1_0.f90: Likewise.
* gfortran.dg/lto/pr79108_0.f90: Likewise.

From-SVN: r264451

[PR87054] fix unaligned access

Building an ADDR_EXPR uses the canonical type to build the pointer
type, but then, as we dereference it, we lose track of lax alignment
known to apply to the dereferenced object.  This might not be a
problem in general, but it is when the compiler implicitly introduces
address taking and dereferencing, as it does for asm statements, and
as it may do in some loop optimizations.

From: Richard Biener <rguenther@suse.de>
for  gcc/ChangeLog

PR middle-end/87054
* gimplify.c (gimplify_expr): Retain alignment of
addressable lvalue in dereference.

From: Alexandre Oliva <oliva@adacore.com>
for  gcc/testsuite/ChangeLog

PR middle-end/87054
* gcc.dg/pr87054.c: New.

From-SVN: r264450

[PR87013] check for .loc is_stmt support in the assembler

Back when we had the logic to output is_stmt but never exercised it,
it didn't matter that we didn't test for assembler support for it.
But there are still assemblers out there that do not support it, so
now that we enable the formerly latent is_stmt logic, we'd better make
sure the assembler can deal with it.

for  gcc/ChangeLog

PR bootstrap/87013
* configure.ac: Check for .loc is_stmt support.
* configure, config.in: Rebuilt.
* dwarf2out.c (dwarf2out_source_line): Skip is_stmt
if not supported.

From-SVN: r264449

gfortran.h (gfc_str_startswith): New macro.

2018-09-20  Janus Weil  <janus@gcc.gnu.org>

* gfortran.h (gfc_str_startswith): New macro.
* decl.c (variable_decl, gfc_match_end): Use it.
* iresolve.c (is_trig_resolved): Ditto.
* module.c (load_omp_udrs, read_module): Ditto.
* options.c (gfc_handle_runtime_check_option): Ditto.
* primary.c (match_arg_list_function): Ditto.
* trans-decl.c (gfc_get_symbol_decl): Ditto.
* trans-expr.c (gfc_conv_procedure_call): Ditto.
* interface.c (dtio_op): Replace strncmp by strcmp.
* resolve.c (resolve_actual_arglist, resolve_function): Ditto.
* trans-expr.c (conv_arglist_function): Ditto.
* trans-intrinsic.c (gfc_conv_ieee_arithmetic_function): Replace macro
STARTS_WITH by gfc_str_startswith.

From-SVN: r264448

rs6000: Remove -misel={yes,no}

These options have been deprecated for many years, supplanted by -misel
and -mno-isel.  This patch finally removes them.

* config/rs6000/rs6000.opt (misel=no, misel=yes): Delete.
* doc/invoke.texi (RS/6000 and PowerPC Options): Delete -misel=yes and
-misel=no.

From-SVN: r264447

[OpenACC] handle missing OMP_LIST_ clauses in fortran's parse tree debugger

gcc/fortran/
* dump-parse-tree.c (show_omp_clauses): Add missing omp list_types
and reorder the switch cases to match the enum in gfortran.h.

From-SVN: r264446

rs6000: Delete VECTOR_OTHER

It's never used.

* config/rs6000/rs6000-opts.h (enum rs6000_vector): Delete
VECTOR_OTHER.
* config/rs6000/rs6000.c (rs6000_debug_vector_unit): Delete
case VECTOR_OTHER.

From-SVN: r264444

hypot-long-double.cc: Skip on arm*.

2018-09-20  Christophe Lyon  <christophe.lyon@linaro.org>

* testsuite/26_numerics/headers/cmath/hypot-long-double.cc: Skip
on arm*.

From-SVN: r264443

PR c++/87075 - ICE with constexpr array initialization.

My patch of 2016-08-26 to avoid calling a trivial default constructor
introduced TARGET_EXPRs initialized with void_node to express trivial
initialization.  But when this shows up in a VEC_INIT_EXPR, we weren't
prepared to handle it.  Fixed by handling it explicitly in
cxx_eval_vec_init_1.

* constexpr.c (cxx_eval_vec_init_1): Handle trivial initialization.

From-SVN: r264442

* doc/invoke.texi: Add -Wno-init-list-lifetime to C++ Language Options.

From-SVN: r264441

Fix PEELING_FOR_NITERS calculation (PR 87288)

PEELING_FOR_GAPS now means "peel one iteration for the epilogue",
in much the same way that PEELING_FOR_ALIGNMENT > 0 means
"peel that number of iterations for the prologue".  We weren't
taking this into account when deciding whether we needed to peel
further scalar iterations beyond the iterations for "gaps" and
"alignment".

Only the first test failed before the patch.  The other two
are just for completeness.

2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR tree-optimization/87288
* tree-vect-loop.c (vect_analyze_loop_2): Take PEELING_FOR_GAPS
into account when determining PEELING_FOR_NITERS.

gcc/testsuite/
PR tree-optimization/87288
* gcc.dg/vect/pr87288-1.c: New test.
* gcc.dg/vect/pr87288-2.c: Likewise,
* gcc.dg/vect/pr87288-3.c: Likewise.

From-SVN: r264440

Add missing alignment checks in epilogue loop vectorisation (PR 86877)

Epilogue loop vectorisation skips vect_enhance_data_refs_alignment
since it doesn't make sense to version or peel the epilogue loop
(that will already have happened for the main loop).  But this means
that it also fails to check whether the accesses are suitably aligned
for the new vector subarch.

We don't seem to carry alignment information from the (potentially
peeled or versioned) main loop to the epilogue loop, which would be
good to fix at some point.  I think we want this patch regardless,
since there's no guarantee that the alignment requirements are the
same for every subarch.

2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR tree-optimization/86877
* tree-vect-loop.c (vect_analyze_loop_2): Call
vect_verify_datarefs_alignment.

gcc/testsuite/
PR tree-optimization/86877
* gfortran.dg/vect/vect-8-epilogue.F90: New test.

From-SVN: r264439

Add -Wclass-conversion.

* c.opt (Wclass-conversion): New.

* decl.c (grok_op_properties): Change a warning from -Wconversion to
-Wclass-conversion.  Make it print the types.

* doc/invoke.texi: Document -Wclass-conversion.

* g++.dg/conversion/op4.C: Add dg-warning.
* g++.dg/warn/Wclass-conversion1.C: New test.
* g++.dg/warn/Wclass-conversion2.C: New test.
* g++.dg/warn/Wconversion5.C: Remove file.
* g++.dg/warn/conversion-function-1.C: Use -Wno-class-converison.
* g++.old-deja/g++.bugs/900215_01.C: Adjust dg-warning.
* g++.old-deja/g++.jason/conversion5.C: Likewise.

From-SVN: r264438

pa.c (pa_adjust_priority): Delete.

* config/pa/pa.c (pa_adjust_priority): Delete.
(TARGET_SCHED_ADJUST_PRIORITY): Delete define.

From-SVN: r264437

pa.md (atomic_storeqi): Restore deleted expander.

* config/pa/pa.md (atomic_storeqi): Restore deleted expander.
(atomic_storehi): Likewise.
(atomic_storesi): Likewise.
(atomic_loaddi): Restore compare and swap exchange loop code.

From-SVN: r264433

Daily bump.

From-SVN: r264432

re PR c++/87324 (g++ ICE with overriding initializers: Segmentation fault)

/cp
2018-09-19  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/87324
* parser.c (cp_parser_initializer_list): Assign error_mark_node
to the index upon error.

/testsuite
2018-09-19  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/87324
* g++.dg/cpp0x/desig5.C: New.

From-SVN: r264428

re PR fortran/84109 (ICE in adjustl on allocatable array of strings)

2018-09-19  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/84109
* trans-expr.c (gfc_trans_assignment_1): The rse.pre for the
assignment of deferred character intrinsic elemental function
results to a realocatable lhs must not be added to the exterior
block if they are array valued but must go to the loop body.

2018-09-19  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/84109
* gfortran.dg/elemental_function_3.f90 : New test.

From-SVN: r264427

combine: Use correct mode in new comparison (PR86902)

This code in try_combine uses the wrong mode.  This fails (with RTL
checking) in trunk, but not in any released branches.

PR rtl-optimization/86902
* combine.c (try_combine): When changing the CC mode used, don't change
an unrelated mode in other_insn to that new CC mode.

From-SVN: r264426

PR c++/87357 - missing -Wconversion warning

PR c++/87357 - missing -Wconversion warning
* decl.c (grok_op_properties): Remove diagnostic parts mentioning
a conversion to a reference to void.  Use
same_type_ignoring_top_level_qualifiers_p rather than comparing types
directly.

* g++.dg/warn/Wconversion5.C: New test.

From-SVN: r264425

dump_printf: use %T and %G throughout

As promised at Cauldron, this patch uses %T and %G with dump_printf and
dump_printf_loc calls to eliminate calls to

  dump_generic_expr (MSG_*, arg, TDF_SLIM)  (via %T)

and

  dump_gimple_stmt (MSG_*, TDF_SLIM, stmt, 0)  (via %G)

throughout the middle-end, simplifying numerous dump callsites.

A few calls to these functions didn't match the above pattern; I didn't
touch these.  I wasn't able to use %E anywhere.

gcc/ChangeLog:
* tree-data-ref.c (runtime_alias_check_p): Use formatted printing
with %T in place of calls to dump_generic_expr.
(prune_runtime_alias_test_list): Likewise.
(create_runtime_alias_checks): Likewise.
* tree-vect-data-refs.c (vect_check_nonzero_value): Likewise.
(vect_analyze_data_ref_dependence): Likewise.
(vect_slp_analyze_data_ref_dependence): Likewise.
(vect_record_base_alignment): Likewise.  Use %G in place of call
to dump_gimple_stmt.
(vect_compute_data_ref_alignment): Likewise.
(verify_data_ref_alignment): Likewise.
(vect_find_same_alignment_drs): Likewise.
(vect_analyze_group_access_1): Likewise.
(vect_analyze_data_ref_accesses): Likewise.
(dependence_distance_ge_vf): Likewise.
(dump_lower_bound): Likewise.
(vect_prune_runtime_alias_test_list): Likewise.
(vect_find_stmt_data_reference): Likewise.
(vect_analyze_data_refs): Likewise.
(vect_create_addr_base_for_vector_ref): Likewise.
(vect_create_data_ref_ptr): Likewise.
* tree-vect-loop-manip.c (vect_set_loop_condition): Likewise.
(vect_can_advance_ivs_p): Likewise.
(vect_update_ivs_after_vectorizer): Likewise.
(vect_gen_prolog_loop_niters): Likewise.
(vect_prepare_for_masked_peels): Likewise.
* tree-vect-loop.c (vect_determine_vf_for_stmt): Likewise.
(vect_determine_vectorization_factor): Likewise.
(vect_is_simple_iv_evolution): Likewise.
(vect_analyze_scalar_cycles_1): Likewise.
(vect_analyze_loop_operations): Likewise.
(report_vect_op): Likewise.
(vect_is_slp_reduction): Likewise.
(check_reduction_path): Likewise.
(vect_is_simple_reduction): Likewise.
(vect_create_epilog_for_reduction): Likewise.
(vect_finalize_reduction:): Likewise.
(vectorizable_induction): Likewise.
(vect_transform_loop_stmt): Likewise.
(vect_transform_loop): Likewise.
(optimize_mask_stores): Likewise.
* tree-vect-patterns.c (vect_pattern_detected): Likewise.
(vect_split_statement): Likewise.
(vect_recog_over_widening_pattern): Likewise.
(vect_recog_average_pattern): Likewise.
(vect_determine_min_output_precision_1): Likewise.
(vect_determine_precisions_from_range): Likewise.
(vect_determine_precisions_from_users): Likewise.
(vect_mark_pattern_stmts): Likewise.
(vect_pattern_recog_1): Likewise.
* tree-vect-slp.c (vect_get_and_check_slp_defs): Likewise.
(vect_record_max_nunits): Likewise.
(vect_build_slp_tree_1): Likewise.
(vect_build_slp_tree_2): Likewise.
(vect_print_slp_tree): Likewise.
(vect_analyze_slp_instance): Likewise.
(vect_detect_hybrid_slp_stmts): Likewise.
(vect_detect_hybrid_slp_1): Likewise.
(vect_slp_analyze_operations): Likewise.
(vect_slp_analyze_bb_1): Likewise.
(vect_transform_slp_perm_load): Likewise.
(vect_schedule_slp_instance): Likewise.
* tree-vect-stmts.c (vect_mark_relevant): Likewise.
(vect_mark_stmts_to_be_vectorized): Likewise.
(vect_init_vector_1): Likewise.
(vect_get_vec_def_for_operand): Likewise.
(vect_finish_stmt_generation_1): Likewise.
(vect_check_load_store_mask): Likewise.
(vectorizable_call): Likewise.
(vectorizable_conversion): Likewise.
(vectorizable_operation): Likewise.
(vectorizable_load): Likewise.
(vect_analyze_stmt): Likewise.
(vect_is_simple_use): Likewise.
(vect_get_vector_types_for_stmt): Likewise.
(vect_get_mask_type_for_stmt): Likewise.
* tree-vectorizer.c (increase_alignment): Likewise.

From-SVN: r264424

Remove constant vec_select restriction.

The vec_select operator is documented to require a const_int for the lane
selector operand, but GCN has an instruction that can select the lane at
runtime, so it seems reasonable to remove this restriction.

This patch simply replaces assertions that the operand is constant with early
exits from the optimizers.  I think it's reasonable that vec_select with a
non-constant operand cannot be optimized, yet.

Also included is the necessary documentation tweak.

2018-09-19  Andrew Stubbs  <ams@codesourcery.com>

gcc/
* doc/rtl.texi: Adjust vec_select description.
* simplify-rtx.c (simplify_binary_operation_1): Allow VEC_SELECT to use
non-constant selectors.

From-SVN: r264423

Don't use __glibcxx_assert to check class invariants

Assertions should be used to check preconditions that users must meet,
not to check whether the implementation is correct.

* include/bits/regex_automaton.tcc (_StateSeq<_TraitsT>::_M_clone()):
Remove __glibcxx_assert statements and use map::find instead of
map::operator[].

From-SVN: r264422

[AARCH64] Use STLUR for atomic_store

Use the STLUR instruction introduced in Armv8.4-a.
This instruction has the store-release semantic like STLR but can take a
9-bit unscaled signed immediate offset.

Example test case:
```
void
foo ()
{
    int32_t *atomic_vals = calloc (4, sizeof (int32_t));
    atomic_store_explicit (atomic_vals + 1, 2, memory_order_release);
}
```

Before patch generates
```
foo:
stp x29, x30, [sp, -16]!
mov x1, 4
mov x0, x1
mov x29, sp
bl calloc
mov w1, 2
add x0, x0, 4
stlr w1, [x0]
ldp x29, x30, [sp], 16
ret
```

After patch generates
```
foo:
stp x29, x30, [sp, -16]!
mov x1, 4
mov x0, x1
mov x29, sp
bl calloc
mov w1, 2
stlur w1, [x0, 4]
ldp x29, x30, [sp], 16
ret
```

We introduce a new feature flag to indicate the presence of this instruction.
The feature flag is called AARCH64_ISA_RCPC8_4 and is included when targeting
armv8.4 architecture.

We also introduce an "arch" attribute to be checked called "rcpc8_4" after this
feature flag.

gcc/

2018-09-19  Matthew Malcomson  <matthew.malcomson@arm.com>

* config/aarch64/aarch64-protos.h
(aarch64_offset_9bit_signed_unscaled_p): New declaration.
* config/aarch64/aarch64.md (arches): New "rcpc8_4" attribute value.
(arch_enabled): Add check for "rcpc8_4" attribute value of "arch".
* config/aarch64/aarch64.h (AARCH64_FL_RCPC8_4): New bitfield.
(AARCH64_FL_FOR_ARCH8_4): Include AARCH64_FL_RCPC8_4.
(AARCH64_FL_PROFILE): Move index so flags are ordered.
(AARCH64_ISA_RCPC8_4): New flag.
* config/aarch64/aarch64.c (offset_9bit_signed_unscaled_p): Renamed
to aarch64_offset_9bit_signed_unscaled_p.
* config/aarch64/atomics.md (atomic_store<mode>): Allow offset
and use stlur.
* config/aarch64/constraints.md (Ust): New constraint.
* config/aarch64/predicates.md.
(aarch64_9bit_offset_memory_operand): New predicate.
(aarch64_rcpc_memory_operand): New predicate.

gcc/testsuite/

2018-09-19  Matthew Malcomson  <matthew.malcomson@arm.com>

* gcc.target/aarch64/atomic-store.c: New.

From-SVN: r264421

re PR rtl-optimization/87361 (gcc.target/sparc/20161111-1.c FAILs)

PR rtl-optimization/87361
* rtlanal.c (nonzero_bits1): Revert accidental change.

From-SVN: r264420

re PR tree-optimization/87349 (ICE: Segmentation fault (in useless_type_conversion_p or is_gimple_reg_type))

2018-09-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87349
PR tree-optimization/87342
* tree-ssa-sccvn.c (do_rpo_vn): Iterate max_rpo computation.

* gcc.dg/torture/pr87349-1.c: New testcase.
* gcc.dg/torture/pr87349-2.c: Likewise.
* gcc.dg/torture/pr87342.c: Likewise.

From-SVN: r264419

libgnuintl.h (_INTL_MAY_RETURN_STRING_ARG, [...]): Backport changes from upstream gettext.

* libgnuintl.h (_INTL_MAY_RETURN_STRING_ARG, gettext, dgettext,
dcgettext, ngettext, dngettext, dcngettext): Backport changes
from upstream gettext.

From-SVN: r264418

Daily bump.

From-SVN: r264417

re PR libstdc++/87135 ([C++17] unordered containers violate iterator validity requirements)

2018-09-18  François Dumont  <fdumont@gcc.gnu.org>

PR libstdc++/87135
* src/c++11/hashtable_c++0x.cc:
(_Prime_rehash_policy::_M_next_bkt): Return a prime no smaller than
requested size, but not necessarily greater.
(_Prime_rehash_policy::_M_need_rehash): Rehash only if target size is
strictly greater than next resize threshold.
* testsuite/23_containers/unordered_map/modifiers/reserve.cc: Adapt test
to validate that there is no rehash as long as number of insertion is
lower or equal to the reserved number of elements.

From-SVN: r264413

re PR fortran/29550 (Optimize -fexternal-blas calls for conjg())

2018-09-18  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/29550
* gfortran.h (gfc_expr): Add external_blas flag.
* frontend-passes.c (matrix_case): Add case A2TB2T.
(optimize_namespace): Handle flag_external_blas by
calling call_external_blas.
(get_array_inq_function): Add argument okind. If
it is nonzero, use it as the kind of argument
to be used.
(inline_limit_check): Remove m_case argument, add
limit argument instead.  Remove assert about m_case.
Set the limit for inlining from the limit argument.
(matmul_lhs_realloc): Handle case A2TB2T.
(inline_matmul_assign): Handle inline limit for other cases with
two rank-two matrices.  Remove no-op calls to inline_limit_check.
(call_external_blas): New function.
* trans-intrinsic.c (gfc_conv_intrinsic_funcall): Do not add
argument to external BLAS if external_blas is already set.

2018-09-18  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/29550
* gfortran.dg/inline_matmul_13.f90: Adjust count for
_gfortran_matmul.
* gfortran.dg/inline_matmul_16.f90: Likewise.
* gfortran.dg/promotion_2.f90: Add -fblas-matmul-limit=1.  Scan
for dgemm instead of dgemm_.  Add call to random_number to make
standard conforming.
* gfortran.dg/matmul_blas_1.f90: New test.
* gfortran.dg/matmul_bounds_14.f: New test.
* gfortran.dg/matmul_bounds_15.f: New test.
* gfortran.dg/matmul_bounds_16.f: New test.
* gfortran.dg/blas_gemm_routines.f: New test / additional file for
preceding tests.

From-SVN: r264412

re PR fortran/29550 (Optimize -fexternal-blas calls for conjg())

2018-09-18  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/29550
* gfortran.h (gfc_expr): Add external_blas flag.
* frontend-passes.c (matrix_case): Add case A2TB2T.
(optimize_namespace): Handle flag_external_blas by
calling call_external_blas.
(get_array_inq_function): Add argument okind. If
it is nonzero, use it as the kind of argument
to be used.
(inline_limit_check): Remove m_case argument, add
limit argument instead.  Remove assert about m_case.
Set the limit for inlining from the limit argument.
(matmul_lhs_realloc): Handle case A2TB2T.
(inline_matmul_assign): Handle inline limit for other cases with
two rank-two matrices.  Remove no-op calls to inline_limit_check.
(call_external_blas): New function.
* trans-intrinsic.c (gfc_conv_intrinsic_funcall): Do not add
argument to external BLAS if external_blas is already set.

2018-09-18  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/29550
* gfortran.dg/inline_matmul_13.f90: Adjust count for
_gfortran_matmul.
* gfortran.dg/inline_matmul_16.f90: Likewise.
* gfortran.dg/promotion_2.f90: Add -fblas-matmul-limit=1.  Scan
for dgemm instead of dgemm_.  Add call to random_number to make
standard conforming.
* gfortran.dg/matmul_blas_1.f90: New test.
* gfortran.dg/matmul_bounds_14.f: New test.
* gfortran.dg/matmul_bounds_15.f: New test.
* gfortran.dg/matmul_bounds_16.f: New test.
* gfortran.dg/blas_gemm_routines.f: New test / additional file for
preceding tests.

From-SVN: r264411

re PR fortran/87239 (ICE in deferred-length string)

2018-09-18  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87239
* trans-expr.c (gfc_trans_assignment_1): The rse.pre for the
assignment of deferred character elemental function results to
a realocatable lhs must not be added to the exterior block but
must go to the loop body.

2018-09-18  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87239
* gfortran.dg/elemental_function_2.f90 : New test.

From-SVN: r264409

P1064R0 - Allowing Virtual Function Calls in Constant Expressions

P1064R0 - Allowing Virtual Function Calls in Constant Expressions
* call.c (build_over_call): No longer check if we're outside a template
function.
* class.c (build_vtbl_initializer): Build vtable's constructor with
indexes.
* constexpr.c (cxx_eval_constant_expression): Don't ignore _vptr's
initializer.  Handle OBJ_TYPE_REF.
(potential_constant_expression_1): Handle OBJ_TYPE_REF.
* decl.c (maybe_commonize_var): Bail out for any DECL_ARTIFICIAL.
(initialize_artificial_var): Mark the variable as constexpr.
(grokdeclarator): Change error to pedwarn.  Only warn when
pedantic and not C++2a.

* gimple-fold.c (gimple_get_virt_method_for_vtable): Adjust assert.

* g++.dg/cpp0x/constexpr-virtual5.C: Adjust dg-error.
* g++.dg/cpp2a/constexpr-virtual1.C: New test.
* g++.dg/cpp2a/constexpr-virtual2.C: New test.
* g++.dg/cpp2a/constexpr-virtual3.C: New test.
* g++.dg/cpp2a/constexpr-virtual4.C: New test.
* g++.dg/cpp2a/constexpr-virtual5.C: New test.
* g++.dg/cpp2a/constexpr-virtual6.C: New test.
* g++.dg/cpp2a/constexpr-virtual7.C: New test.
* g++.dg/cpp2a/constexpr-virtual8.C: New test.
* g++.dg/cpp2a/constexpr-virtual9.C: New test.
* g++.dg/diagnostic/virtual-constexpr.C: Skip for C++2a.  Use
-pedantic-errors.  Adjust dg-error.

From-SVN: r264408

re PR fortran/87336 (wrong output for pointer dummy assiocated to target actual argument)

2018-09-18  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87336
* trans-array.c (gfc_get_array_span): Try to get the element
length of incomplete types. Return NULL_TREE otherwise.
(gfc_conv_expr_descriptor): Only set the 'span' field if the
above does not return NULL_TREE. Set 'span' field if possible
for all new descriptors.

2018-09-18  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87336
* gfortran.dg/pointer_array_10.f90 : New test.
* gfortran.dg/assign_10.f90 : Increase 'parm' count to 20.
* gfortran.dg/transpose_optimization_2.f90 : Increase 'parm'
count to 72.

From-SVN: r264405

rs6000: Remove old "Cygnus sibcall" comment

This comment is quite cryptic and very out-of-date by now.  Committing.

* config/rs6000/rs6000.md: Remove old "Cygnus sibcall" comment.

From-SVN: r264403

re PR c++/85065 ([concepts] ICE with invalid use of a concept)

/cp
2018-09-18  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/85065
* cp-tree.h (NON_ERROR): New.
* pt.c (auto_hash::hash): Use it.
(do_auto_deduction): Likewise.

/testsuite
2018-09-18  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/85065
* g++.dg/concepts/pr85065.C: New.

From-SVN: r264402

Handle CLOBBER in reg_overlap_mentioned_p (PR86882)

Combine will put CLOBBER (with a non-void mode) anywhere in a pattern
to poison it.  reg_overlap_mentioned_p did not handle this.  This patch
fixes that.

PR rtl-optimization/86882
* rtlanal.c (reg_overlap_mentioned_p): Handle CLOBBER.

From-SVN: r264400