Allow inner-loop reductions with variable-length vectors

While working on PR 86871, I noticed we were being overly restrictive
when handling variable-length vectors.  For:

  for (i : ...)
    {
      res = ...;
      for (j : ...)
        res op= ...;
      a[i] = res;
    }

we don't need a reduction operation (although we do for double
reductions like:

  res = ...;
  for (i : ...)
    for (j : ...)
      res op= ...;
  a[i] = res;

which must still be rejected).

2018-08-08  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-loop.c (vectorizable_reduction): Allow inner-loop
reductions for variable-length vectors.

gcc/testsuite/
* gcc.target/aarch64/sve/reduc_8.c: New test.

From-SVN: r263451

diagnostics: add line numbers to source (PR other/84889)

This patch adds a left margin to the lines of source (and annotations)
printed by diagnostic_show_locus, so that e.g. rather than:

test.c: In function 'test':
test.c:12:15: error: 'struct foo' has no member named 'm_bar'; did you mean 'bar'?
   return ptr->m_bar;
               ^~~~~
               bar

we print:

test.c: In function 'test':
test.c:12:15: error: 'struct foo' has no member named 'm_bar'; did you mean 'bar'?
12 |   return ptr->m_bar;
   |               ^~~~~
   |               bar

Similarly, for a multiline case (in C++ this time), this:

bad-binary-ops.C: In function 'int test_2()':
bad-binary-ops.C:26:4: error: no match for 'operator+' (operand types are 's' and 't')
   return (some_function ()
           ~~~~~~~~~~~~~~~~
    + some_other_function ());
    ^~~~~~~~~~~~~~~~~~~~~~~~

becomes:

bad-binary-ops.C: In function 'int test_2()':
bad-binary-ops.C:26:4: error: no match for 'operator+' (operand types are 's' and 't')
25 |   return (some_function ()
   |           ~~~~~~~~~~~~~~~~
26 |    + some_other_function ());
   |    ^~~~~~~~~~~~~~~~~~~~~~~~

I believe this slightly improves the readability of the output, in that it:
- distinguishes between the user's source code vs the annotation lines
  that we're adding (the underlinings and fix-it hints here)
- shows the line numbers in another place (potentially helpful for
  multiline diagnostics, where the user can see the line numbers directly,
  rather than have to figure them out relative to the caret: in the 2nd
  example, note how the diagnostic is reported at line 26, but the first
  line printed is actually line 25)

I'm not sure that this is the precise format we want to go with [1], but
I think it's an improvement over the status quo, and we're in stage 1
of gcc 9, so there's plenty of time to shake out issues.

I've turned it on by default; it can be disabled via
-fno-diagnostics-show-line-numbers (it's also turned off in the testsuite, to
avoid breaking numerous existing test cases).

[1] Some possible variants:
  - maybe just "LL|" rather than "LL | "
  - maybe ':' rather than '|'
  - maybe we should have some leading indentation, to better split up
    the diagnostics visually via the left-hand column
  - etc

gcc/ChangeLog:
PR other/84889
* common.opt (fdiagnostics-show-line-numbers): New option.
* diagnostic-show-locus.c (class layout): Add fields
"m_show_line_numbers_p" and "m_linenum_width";
(num_digits): New function.
(test_num_digits): New function.
(layout::layout): Initialize new fields.  Update m_x_offset
logic to handle any left margin.
(layout::print_source_line): Print line number when requested.
(layout::start_annotation_line): New member function.
(layout::print_annotation_line): Call it.
(layout::print_leading_fixits): Likewise.
(layout::print_trailing_fixits): Likewise.  Update calls to
move_to_column for new parameter.
(layout::get_x_bound_for_row): Add "add_left_margin" param and use
it to potentially call start_annotation_line.
(layout::show_ruler): Call start_annotation_line.
(selftest::test_line_numbers_multiline_range): New selftest.
(selftest::diagnostic_show_locus_c_tests): Call test_num_digits
and selftest::test_line_numbers_multiline_range.
* diagnostic.c (diagnostic_initialize): Initialize
show_line_numbers_p.
* diagnostic.h (struct diagnostic_context): Add field
"show_line_numbers_p".
* doc/invoke.texi (Diagnostic Message Formatting Options): Add
-fno-diagnostics-show-line-numbers.
* dwarf2out.c (gen_producer_string): Add
OPT_fdiagnostics_show_line_numbers to the ignored options.
* lto-wrapper.c (merge_and_complain): Likewise to the "pick
one setting" options.
(append_compiler_options): Likewise to the dropped options.
(append_diag_options): Likewise to the passed-on options.
* opts.c (common_handle_option): Handle the new option.
* toplev.c (general_init): Set up global_dc->show_line_numbers_p.

gcc/testsuite/ChangeLog:
PR other/84889
* gcc.dg/plugin/diagnostic-test-show-locus-bw-line-numbers.c: New
test.
* gcc.dg/plugin/diagnostic-test-show-locus-color-line-numbers.c:
New test.
* gcc.dg/plugin/plugin.exp (plugin_test_list): Add the new tests.
* lib/prune.exp: Add -fno-diagnostics-show-line-numbers to
TEST_ALWAYS_FLAGS.

From-SVN: r263450

extend.texi (PowerPC AltiVec Built-in Functions Available on ISA 2.07): Correct spelling of bcdsub to be __builtin_bcdsub.

gcc/ChangeLog:

2018-08-09  Kelvin Nilsen  <kelvin@gcc.gnu.org>

* doc/extend.texi (PowerPC AltiVec Built-in Functions Available on
ISA 2.07): Correct spelling of bcdsub to be __builtin_bcdsub.  Add
third argument of type "const signed char" to descriptions of
__builtin_bcdadd, __builtin_bcdadd_lt, __builtin_bcdadd_eq,
__builtin_bcdadd_gt, __builtin_bcdadd_ov, __builtin_bcdsub,
__builtin_bcdsub_lt, __builtin_bcdsub_eq, __builtin_bcdsub_gt,
__builtin_bcdsub_ov functions.

From-SVN: r263449

Restore flow_bb_inside_loop_p tests (PR 86858)

The series to remove vinfo_for_stmt also removed tests of
flow_bb_inside_loop_p if the call was simply testing whether the
statement was in the vectorisation region.  I'd tried to keep calls
that were testing whether the statement was in a particular loop
(inner or outer), but messed up in vect_is_simple_reduction and
removed calls that were actually needed.  This patch restores them.

I double-checked the other removed calls and I think these are
the only ones affected.

2018-08-08  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR tree-optimization/86858
* tree-vect-loop.c (vect_is_simple_reduction): Restore
flow_bb_inside_loop_p calls.

gcc/testsuite/
PR tree-optimization/86858
* gcc.dg/vect/pr86858.c: New test.

From-SVN: r263448

Fix invalid assumption in vect_transform_stmt (PR 86871)

The handling of outer-loop uses of inner-loop definitions assumed
that anything that wasn't a PHI would be a gassign.  It's also
possible for it to be a gcall.

2018-08-08  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR tree-optimization/86871
* tree-vect-stmts.c (vect_transform_stmt): Use gimple_get_lhs
instead of gimple_assign_lhs.

gcc/testsuite/
PR tree-optimization/86871
* gcc.dg/vect/pr86871.c: New test.

From-SVN: r263447

aarch64 - PR target/86887 Fix missing register constraints in carryin patterns

Some of the carryin insn patterns are missing a register constraint.
That means that the register allocator can pick practically anything
to hold that value, including memory locations, or registers of the
wrong class.

PR target/86887
* config/aarch64/aarch64.md (add<mode>3_carryinC_zero): Add missing
register constraint to operand 0.
(add<mode>3_carryinC): Likewise.
(add<mode>3_carryinV_zero, add<mode>3_carryinV): Likewise.

From-SVN: r263446

p784.C: Add -fpermissive.

2018-08-09  Paolo Carlini  <paolo.carlini@oracle.com>

* g++.old-deja/g++.mike/p784.C: Add -fpermissive.

From-SVN: r263445

Remove extra line in common.opt (PR c/86895).

2018-08-09  Martin Liska  <mliska@suse.cz>

        PR c/86895
* common.opt: Remove extra line.

From-SVN: r263444

S/390: Fix target attribute tests

These broke after removing G5 and G6 CPUs from the processor
enumeration.

gcc/testsuite/ChangeLog:

2018-08-09  Andreas Krebbel  <krebbel@linux.ibm.com>

* gcc.target/s390/target-attribute/tattr-arch-tune-1.c: Decrement
arch and tune numbers by 2.
* gcc.target/s390/target-attribute/tattr-arch-tune-2.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-1.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-10.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-11.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-12.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-13.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-14.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-15.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-16.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-17.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-18.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-19.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-2.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-20.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-21.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-22.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-23.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-24.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-25.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-26.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-27.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-28.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-29.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-3.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-30.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-31.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-32.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-4.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-5.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-6.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-7.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-8.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m31-9.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-1.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-10.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-11.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-12.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-13.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-14.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-15.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-16.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-17.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-18.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-19.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-2.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-20.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-21.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-22.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-23.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-24.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-25.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-26.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-27.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-28.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-29.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-3.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-30.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-31.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-32.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-33.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-4.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-5.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-6.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-7.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-8.c: Likewise.
* gcc.target/s390/target-attribute/tattr-m64-9.c: Likewise.

From-SVN: r263443

Fix typos in params.def.

2018-08-09  Martin Liska  <mliska@suse.cz>

* params.def (PARAM_ALIGN_LOOP_ITERATIONS): Remove double dots
        at the end of a line, make first letter capital and end up
        a sentence with a dot.
(PARAM_LOOP_INTERCHANGE_STRIDE_RATIO): Likewise.
(PARAM_LOOP_BLOCK_TILE_SIZE): Likewise.
(PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS): Likewise.
(PARAM_GRAPHITE_MAX_ARRAYS_PER_SCOP): Likewise.
(PARAM_MAX_ISL_OPERATIONS): Likewise.
(PARAM_GRAPHITE_ALLOW_CODEGEN_ERRORS): Likewise.
(PARAM_PROFILE_FUNC_INTERNAL_ID): Likewise.
(PARAM_INDIR_CALL_TOPN_PROFILE): Likewise.
(PARAM_SLP_MAX_INSNS_IN_BB): Likewise.
(PARAM_IPA_CP_EVAL_THRESHOLD): Likewise.
(PARAM_IPA_CP_RECURSION_PENALTY): Likewise.
(PARAM_IPA_CP_SINGLE_CALL_PENALTY): Likewise.
(PARAM_IPA_CP_LOOP_HINT_BONUS): Likewise.
(PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS): Likewise.
(PARAM_TREE_REASSOC_WIDTH): Likewise.
(PARAM_HSA_GEN_DEBUG_STORES): Likewise.
(PARAM_MAX_SPECULATIVE_DEVIRT_MAYDEFS): Likewise.
(PARAM_MAX_VRP_SWITCH_ASSERTIONS): Likewise.

From-SVN: r263442

S/390: Fix PR84332 ICE with stack clash protection

Our implementation of the stack probe requires the probe interval to
be used as displacement in an address operand.  The maximum probe
interval currently is 64k.  This would exceed short displacements.
Trim that value down to 4k if that happens.  This might result in too
many probes being generated only on the oldest supported machine level
z900.

gcc/ChangeLog:

2018-08-09  Andreas Krebbel  <krebbel@linux.ibm.com>

PR target/84332
* config/s390/s390.c (s390_option_override_internal): Reduce the
stack-clash-protection-probe-interval param if it would be too big
for z900.

gcc/testsuite/ChangeLog:

2018-08-09  Andreas Krebbel  <krebbel@linux.ibm.com>

PR target/84332
* gcc.target/s390/pr84332.c: New testcase.

From-SVN: r263441

Daily bump.

From-SVN: r263439

2018-08-08  François Dumont  <fdumont@gcc.gnu.org>

* include/bits/stl_algo.h
(__rotate(_Ite, _Ite, _Ite, forward_iterator_tag))
(__rotate(_Ite, _Ite, _Ite, bidirectional_iterator_tag))
(__rotate(_Ite, _Ite, _Ite, random_access_iterator_tag)): Move code duplication...
(rotate(_Ite, _Ite, _Ite)): ...here.
(__stable_partition_adaptive(_FIt, _FIt, _Pred, _Dist, _Pointer, _Dist)):
Simplify rotate call.
(__rotate_adaptive(_BIt1, _BIt1, _BIt1, _Dist, _Dist, _Bit2, _Dist)):
Likewise.
(__merge_without_buffer(_BIt, _BIt, _BIt, _Dist, _Dist, _Comp)):
Likewise.

From-SVN: r263433

m68k: handle more cases of TLS symbols with offset

PR target/46179
* config/m68k/m68k.h (FINAL_PRESCAN_INSN): Don't define.
* config/m68k/m68k.c (handle_move_double): Don't call
m68k_final_prescan_insn.
(m68k_adjust_decorated_operand): Renamed from
m68k_final_prescan_insn, remove first and third operand and
simplify.
(print_operand): Call it.
(print_operand_address): Call it.

PR target/46179
* gcc.target/m68k/tls-dimode.c: New file.

From-SVN: r263432

[PATCH] line-map include-from representation

https://gcc.gnu.org/ml/gcc-patches/2018-08/msg00554.html
Make linemap::included_from a location
libcpp/
* include/line-map.h (struct line_map_ordinary): Replace
included_from map index with included_at source_location.
(ORDINARY_MAP_INCLUDER_FILE_INDEX): Delete.
(LAST_SOURCE_LINE_LOCATION): Delete.
(LAST_SOURCE_LINE, LAST_SOURCE_COLUMN): Delete.
(linemap_included_from): New.
(linemap_included_from_linemap): Declare.
(MAIN_FILE_P): Adjust.
* line-map.c (linemap_included_from_linemap): New.
(lonemap_check_files_exited): Use linemap_included_at.
(linemap_add): Adjust inclusion setting.
(linemap_dump, linemap_dump_location): Adjust.
* directives.c (do_linemarker): Use linemap_included_at.
gcc/
* diagnostic.c (diagnostic_report_current_module): Use
linemap_included_from & linemap_included_from_linemap.
gcc/c-family/
* c-common.c (try_to_locate_new_include_inertion_point): Use
linemap_included_from_linemap.
* c-lex.c (fe_file_change): Use linemap_included_from.
* c-ppoutput.c (pp_file_change): Likewise.
gcc/fortran/
* cpp.c (cb_file_change): Use linemap_included_from.
gcc/testsuite/
* c-c++-common/inc-from-1a.h, c-c++-common/inc-from-1b.h,
c-c++-common/inc-from-1.c: New

From-SVN: r263429

Prevent internal aligned_alloc clashing with libc version

If configure fails to detect aligned_alloc we will try to define our
own in new_opa.cc but that could clash with the libcversion in
<stdlib.h>. Use a namespace to keep them distinct.

* libsupc++/new_opa.cc (aligned_alloc): Declare inside namespace to
avoid clashing with an ::aligned_alloc function that was not detected
by configure.

From-SVN: r263409

[libgomp, nvptx] Fall back to cuLinkAddData/cuLinkCreate if _v2 not found

Cuda driver api functions cuLinkAddData and cuLinkCreate are available starting
version 5.5.  In version 6.5, they are remapped onto _v2 versions.

The dlopen interface of the libgomp nvptx plugin uses the _v2 versions, so it
won't work with a cuda driver with driver api version lower than 6.5.

This patch fixes the problem by testing for the presence of the _v2 versions,
and falling back to the original versions in case of absence of the _v2
versions.

Build on x86_64 with nvptx accelerator and reg-tested libgomp, both with and
without --without-cuda-driver.

2018-08-08  Tom de Vries  <tdevries@suse.de>

* plugin/cuda-lib.def (cuLinkAddData_v2, cuLinkCreate_v2): Declare using
CUDA_ONE_CALL_MAYBE_NULL.
* plugin/plugin-nvptx.c (cuLinkAddData, cuLinkCreate): Undef and declare.
(cuLinkAddData_v2, cuLinkCreate_v2): Declare.
(link_ptx): Fall back to cuLinkAddData/cuLinkCreate if the _v2 versions
are not found.

From-SVN: r263408

[libgomp, nvptx] Allow cuGetErrorString to be NULL

Cuda driver api function cuGetErrorString is available in version 6.0 and
higher.

Currently, when the driver that is used does not contain this function, the
libgomp nvptx plugin will not build (PLUGIN_NVPTX_DYNAMIC == 0) or run
(PLUGIN_NVPTX_DYNAMIC == 1).

This patch fixes this problem by testing for the presence of the function, and
handling absence.

Build on x86_64 with nvptx accelerator and reg-tested libgomp, both with and
without --without-cuda-driver.

2018-08-08  Tom de Vries  <tdevries@suse.de>

* plugin/cuda-lib.def (cuGetErrorString): Use CUDA_ONE_CALL_MAYBE_NULL.
* plugin/plugin-nvptx.c (cuda_error): Handle if cuGetErrorString is not
present.

From-SVN: r263407

[libgomp, nvptx] Remove hard-coded const in nvptx_open_device

CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR is defined in cuda driver
api version 6.0 and higher.

Currently nvptx_open_device uses a hard-coded constant instead.

This patch fixes that by:
- defining CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR to the hardcoded
  constant at toplevel, if not present in cuda.h, and
- using CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR in nvptx_open_device

Build on x86_64 with nvptx accelerator and reg-tested libgomp.

2018-08-08  Tom de Vries  <tdevries@suse.de>

* plugin/plugin-nvptx.c
(CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR): Define.
(nvptx_open_device): Use
CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR.

From-SVN: r263406

[libgomp, nvptx] Note that cuGetErrorString is in CUDA_VERSION >= 6000

Cuda driver api function cuGetErrorString is available in version 6.0 and
higher.

This patch:
- removes a comment saying the declaration is not available in cuda.h 6.0
- fixes the presence test to use CUDA_VERSION < 6000
- moves the declaration to toplevel

Build on x86_64 with nvptx accelerator and reg-tested libgomp.

2018-08-08  Tom de Vries  <tdevries@suse.de>

* plugin/plugin-nvptx.c (cuda_error): Move declaration of cuGetErrorString ...
(cuGetErrorString): ... here.  Guard with CUDA_VERSION < 6000.

From-SVN: r263405

[AArch64] Add HXT Phecda core support

HXT semiconductor's CPU core Phecda, as a variant of Qualcomm qdf24xx,
reuses the same tuning structure and pipeline with it.

Applied on behalf of: Hongbo Zhang <hongbo.zhang@linaro.org>

* config/aarch64/aarch64-cores.def: Add phecda core.
* config/aarch64/aarch64-tune.md: Regenerate.
* doc/invoke.texi: Add phecda core.

From-SVN: r263404

Fix Docbook markup for table entry

* doc/xml/manual/using.xml: Fix markup for empty table entry.
* doc/html/*: Regenerate.

From-SVN: r263401

Add missing <experimental/numeric> header to docs

* doc/xml/manual/using.xml: Add missing header to table and fix typo.
* doc/html/*: Regenerate.

From-SVN: r263398

PR libstdc++/86597 directory_entry observers should clear error_code

PR libstdc++/86597
* include/bits/fs_dir.h (directory_entry::_M_file_type(error_code&)):
Clear error_code when cached type is used.
* testsuite/27_io/filesystem/directory_entry/86597.cc: New test.

From-SVN: r263397

S/390: Fix PR85295

gcc/ChangeLog:

2018-08-08  Andreas Krebbel  <krebbel@linux.ibm.com>

PR target/85295
* config/s390/constraints.md ("NxHD0", "NxSD0"): New constraint
definitions.
* config/s390/s390.md ("movti"): Add more alternatives for
constant to GPR copies.

gcc/testsuite/ChangeLog:

2018-08-08  Andreas Krebbel  <krebbel@linux.ibm.com>

PR target/85295
* gcc.target/s390/TI-constants-lra.c: New testcase.
* gcc.target/s390/TI-constants-nolra.c: New testcase.

From-SVN: r263396

S/390: Fix whitespace problems in the backend

2018-08-08  Andreas Krebbel  <krebbel@linux.ibm.com>

* config/s390/s390.c: Fix whitespace damage throughout the file.
* config/s390/s390.h: Likewise.
* config/s390/tpf.h: Likewise.

From-SVN: r263395

S/390: Remove TARGET_CPU_ZARCH

TARGET_CPU_ZARCH allowed to distinguish between g5/g6 and newer
machines.  Since the former are now gone, we can assume that
TARGET_CPU_ZARCH is always true.  As a side-effect, branch splitting
is now completely gone.  Some parts of literal pool splitting are also
gone, but it's still there: we need to support it because floating
point and vector instructions still cannot use relative addressing.

gcc/ChangeLog:

2018-08-08  Ilya Leoshkevich  <iii@linux.ibm.com>

* config/s390/s390.c (s390_loadrelative_operand_p):
        Remove TARGET_CPU_ZARCH usages.
(s390_rtx_costs): Likewise.
(s390_legitimate_constant_p): Likewise.
(s390_cannot_force_const_mem): Likewise.
(legitimate_reload_constant_p): Likewise.
(s390_preferred_reload_class): Likewise.
(legitimize_pic_address): Likewise.
(legitimize_tls_address): Likewise.
(s390_split_branches): Removed.
(s390_add_execute): Removed.
(s390_dump_pool): Remove TARGET_CPU_ZARCH usages.
(s390_mainpool_start): Likewise.
(s390_mainpool_finish): Likewise.
(s390_mainpool_cancel): Removed.
(s390_chunkify_start): Remove TARGET_CPU_ZARCH usages.
(s390_chunkify_cancel): Likewise.
(s390_return_addr_rtx): Likewise.
(s390_register_info): Remove split_branches_pending_p uages.
(s390_optimize_register_info): Likewise.
(s390_init_frame_layout): Remove TARGET_CPU_ZARCH and
        split_branches_pending_p usages.
(s390_can_eliminate): Remove TARGET_CPU_ZARCH usages.
(s390_load_got): Likewise.
(s390_expand_split_stack_prologue): Likewise.
(output_asm_nops): Likewise.
(s390_function_profiler): Likewise.
(s390_emit_call): Likewise.
(s390_conditional_register_usage): Likewise.
(s390_optimize_prologue): Likewise.
(s390_reorg): Remove TARGET_CPU_ZARCH and
        split_branches_pending_p usages.
(s390_option_override_internal): Remove TARGET_CPU_ZARCH
        usages.
(s390_output_indirect_thunk_function): Likewise.
* config/s390/s390.h (TARGET_CPU_ZARCH): Removed.
(TARGET_CPU_ZARCH_P): Removed.
(struct machine_function): Remove split_branches_pending_p.
* config/s390/s390.md: Remove TARGET_CPU_ZARCH usages.

From-SVN: r263394

S/390: Remove support for g5 and g6 machines

g5 and g6 were deprecated since gcc 6.1.0 (commit 3bd8520f).

gcc/ChangeLog:

2018-08-08  Ilya Leoshkevich  <iii@linux.ibm.com>

* common/config/s390/s390-common.c (processor_flags_table):
        Remove flags.
* config.gcc: Remove with_arch/with_tune support.
* config/s390/2064.md: Remove cpu attribute comparisons.
* config/s390/driver-native.c (s390_host_detect_local_cpu):
        Remove MTN.
* config/s390/linux.h (ASM_SPEC):
        Remove -march support.
* config/s390/s390-c.c (s390_cpu_cpp_builtins_internal):
        Use a table to get an arch level.
* config/s390/s390-opts.h (enum processor_type):
        Remove enum values.
* config/s390/s390.c
        (processor_table): Remove entries, add arch_level values.
        (s390_issue_rate): Remove cases.
        (s390_option_override): Adjust
        s390_option_override_internal() call.
(s390_option_override_internal): Remove deprecation warning.
        (s390_valid_target_attribute_tree): Adjust
        s390_option_override_internal() call.
* config/s390/s390.h (struct s390_processor):
        Share with s390-c.c, add arch_level field.
* config/s390/s390.md:
        Remove occurrences in cpu attribute.
* config/s390/s390.opt: Remove -march/-mtune support.
* config/s390/tpf.h (ASM_SPEC): Remove -march support.
* doc/invoke.texi: Remove deprecation warning.

gcc/testsuite/ChangeLog:

2018-08-08  Ilya Leoshkevich  <iii@linux.ibm.com>

* gcc.target/s390/hotpatch-8.c: Remove.
* gcc.target/s390/hotpatch-9.c: Remove.
* gcc.target/s390/mnop-mcount-m31-fpic.c: Remove.
* gcc.target/s390/mnop-mcount-m31.c: Remove.

From-SVN: r263393

P0595R1 - is_constant_evaluated

P0595R1 - is_constant_evaluated
cp/
* cp-tree.h (enum cp_built_in_function): New.
(maybe_constant_init): Add pretend_const_required argument.
* typeck2.c (store_init_value): Pass true as new argument to
maybe_constant_init.
* constexpr.c (constexpr_fn_retval): Check also DECL_BUILT_IN_CLASS
for BUILT_IN_UNREACHABLE.
(struct constexpr_ctx): Add pretend_const_required field.
(cxx_eval_builtin_function_call): Use DECL_IS_BUILTIN_CONSTANT_P
macro.  Handle CP_BUILT_IN_IS_CONSTANT_EVALUATED.  Check also
DECL_BUILT_IN_CLASS for BUILT_IN_UNREACHABLE.
(cxx_eval_outermost_constant_expr): Add pretend_const_required
argument, initialize pretend_const_required field in ctx.  If the
result is TREE_CONSTANT and non_constant_p, retry with
pretend_const_required false if it was true.
(is_sub_constant_expr): Initialize pretend_const_required_field in
ctx.
(cxx_constant_value): Pass true as pretend_const_required to
cxx_eval_outermost_constant_expr.
(maybe_constant_value): Pass false as pretend_const_required to
cxx_eval_outermost_constant_expr.
(fold_non_dependent_expr): Likewise.
(maybe_constant_init_1): Add pretend_const_required argument, pass it
down to cxx_eval_outermost_constant_expr.  Pass !allow_non_constant
instead of false as strict to cxx_eval_outermost_constant_expr.
(maybe_constant_init): Add pretend_const_required argument, pass it
down to maybe_constant_init_1.
(cxx_constant_init): Pass true as pretend_const_required to
maybe_constant_init_1.
* cp-gimplify.c (cp_gimplify_expr): Handle CALL_EXPRs to
CP_BUILT_IN_IS_CONSTANT_EVALUATED.
(cp_fold): Don't fold CP_BUILT_IN_IS_CONSTANT_EVALUATED calls.
* decl.c: Include langhooks.h.
(cxx_init_decl_processing): Register __builtin_is_constant_evaluated
built-in.
* tree.c (builtin_valid_in_constant_expr_p): Return true for
CP_BUILT_IN_IS_CONSTANT_EVALUATED.
* pt.c (declare_integer_pack): Initialize DECL_FUNCTION_CODE.
testsuite/
* g++.dg/cpp2a/is-constant-evaluated1.C: New test.

From-SVN: r263392

re PR c++/86836 (internal compiler error on structured bindings with shadow parameter on templated function)

PR c++/86836
* pt.c (tsubst_expr): For structured bindings, call tsubst_decomp_names
before tsubst_init, not after it.

* g++.dg/cpp1z/decomp46.C: New test.

From-SVN: r263391

re PR c++/86738 (gcc 8.2: Internal compiler error memcpy)

PR c++/86738
* constexpr.c (cxx_eval_binary_expression): For arithmetics involving
NULL pointer set *non_constant_p to true.
(cxx_eval_component_reference): For dereferencing of a NULL pointer,
set *non_constant_p to true and return t.

* g++.dg/opt/pr86738.C: New test.

From-SVN: r263390

[AArch64, Falkor] Switch to using Falkor-specific vector costs.

The adjusted vector costs give Falkor a reasonable boost in performance for FP
benchmarks (both CPU2017 and CPU2006) and doesn't change INT benchmarks that
much. There are some regressions that will be investigated as follow on work.

Numbers from the CI run:

CPU2017:

(R) 605.mcf_s: -1.8%
(R) 620.omnetpp_s: -2%
    623.xalancbmk_s: 2%
    654.roms_s: 7%

(R) INT mean: -0.09%
    FP mean: 0.70%

CPU2006:

(R) 429.mc: -5%
(R) 471.omnetpp: -9.5% (potentially noise/fluctuations)
    483.xalancbmk: 6.02%
    410.bwaves: 5.03%
    433.milc: 2%
    434.zeusmp: 10.5%
(R) 436.cactusADM: -12.75%
    437.leslie3d: 5.94%
(R) 453.povray: -0.82%
    459.GemsFDTD: 16.87%
    465.tonto: 1%

(R) INT mean: -0.79%
    FP mean: 1.54%

gcc/ChangeLog:

2018-08-08  Luis Machado  <luis.machado@linaro.org>

* config/aarch64/aarch64.c (qdf24xx_vector_cost): New static global.
(qdf24xx_tunings): Set vector cost structure to qdf24xx_vector_cost.

From-SVN: r263389

[aarch64] Adjust Falkor's sign extend reg+reg address cost

Adjust Falkor's register_sextend cost from 4 to 3.  This fixes a testsuite
failure in gcc.target/aarch64/extend.c:ldr_sxtw where GCC was generating
a sbfiz instruction rather than a load with sign extension.

No performance changes.

gcc/ChangeLog:

2018-08-08  Luis Machado  <luis.machado@linaro.org>

* config/aarch64/aarch64.c (qdf24xx_addrcost_table)
<register_sextend>: Set to 3.

From-SVN: r263388

Daily bump.

From-SVN: r263387

[OpenACC] cleanup trans-stmt.h

gcc/fortran/
* trans-stmt.h: Remove stale reference to trans-openacc.c.

From-SVN: r263372

[PATCH] Line map table allocation

https://gcc.gnu.org/ml/gcc-patches/2018-08/msg00434.html
* line-map.c: (linemap_init): Set default allocator here.
(new_linemap): Rather than here.  Refactor allocation logic.

From-SVN: r263366

PR libstdc++/86874 fix std::variant::swap regression

PR libstdc++/86874
* include/std/variant (_Copy_ctor_base::_M_destructive_move): Define
here instead of in _Move_assign_base.
(_Copy_ctor_base<true, _Types...>::_M_destructive_move): Define.
(_Copy_assign_base::operator=): Use _M_destructive_move when changing
the contained value to another alternative.
(_Move_assign_base::operator=): Likewise.
(_Move_assign_base::_M_destructive_move): Remove.
* testsuite/20_util/variant/86874.cc: New test.

From-SVN: r263365

runtime: use poll rather than pollset for netpoll on AIX

    Updates golang/go#26634

    Reviewed-on: https://go-review.googlesource.com/126857

From-SVN: r263364

libgo: uncomment trace.Stop() call in testing package

    Fix up the testing package to insure that execution traces
    work properly (e.g. "-test.trace=<XXX>" command line option). The
    call to stop tracing and emit the output file was stubbed out.

    Reviewed-on: https://go-review.googlesource.com/128275

From-SVN: r263363

[AArch64] Fix -mlow-precision-div (PR 86838)

The "@" handling broke -mlow-precision-div, because the scalar forms of
the instruction were provided by a pattern that also provided FRECPX
(and so were parameterised on an unspec code as well as a mode),
while the SIMD versions had a dedicated FRECPE pattern.  This patch
moves the scalar FRECPE handling to the SIMD pattern too (as for FRECPS)
and uses a separate pattern for FRECPX.

The convention in aarch64-simd-builtins.def seemed to be to add
comments only if the mapping wasn't obvious (i.e. not just sticking
"aarch64_" on the beginning and "<mode>" on the end), so the patch
deletes the reference to the combined pattern instead of rewording it.

There didn't seem to be any coverage of -mlow-precision-div in the
testsuite, so the patch adds some tests for it.

2018-08-07  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR target/86838
* config/aarch64/iterators.md (FRECP, frecp_suffix): Delete.
* config/aarch64/aarch64-simd.md
(aarch64_frecp<FRECP:frecp_suffix><mode>): Fold FRECPE into...
(@aarch64_frecpe<mode>): ...here and the move FRECPX to...
(aarch64_frecpx<mode>): ...this new pattern.
* config/aarch64/aarch64-simd-builtins.def: Remove comment
about aarch64_frecp<FRECP:frecp_suffix><mode>.

gcc/testsuite/
PR target/86838
* gcc.target/aarch64/frecpe_1.c: New test.
* gcc.target/aarch64/frecpe_2.c: Likewise.

From-SVN: r263362

PR c++/59480, DR 136

/cp
2018-08-07  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/59480, DR 136
* decl.c (check_no_redeclaration_friend_default_args): New.
(duplicate_decls): Use the latter; also check that a friend
declaration specifying default arguments is a definition.

/testsuite
2018-08-07  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/59480, DR 136
* g++.dg/other/friend8.C: New.
* g++.dg/other/friend9.C: Likewise.
* g++.dg/other/friend10.C: Likewise.
* g++.dg/other/friend11.C: Likewise.
* g++.dg/other/friend12.C: Likewise.
* g++.dg/other/friend13.C: Likewise.
* g++.dg/other/friend14.C: Likewise.
* g++.dg/other/friend15.C: Likewise.
* g++.dg/parse/defarg4.C: Compile with -fpermissive -w.
* g++.dg/parse/defarg8.C: Likewise.

From-SVN: r263361

PR libstdc++/86861 Meet precondition for Solaris memalign

Solaris memalign requires alignment to be at least sizeof(int), so
increase it as needed.

Also move the check for valid alignments from the fallback
implementation of aligned_alloc into operator new, as it's required for
all of aligned_alloc, memalign, posix_memalign and __aligned_malloc.
This adds a branch to check for undefined behaviour which we could just
ignore, so the check could just be removed. It should certainly be
removed if PR 86878 is implemented to issue a warning about calls with
invalid alignments.

PR libstdc++/86861
* libsupc++/new_opa.cc [_GLIBCXX_HAVE_MEMALIGN] (aligned_alloc):
Replace macro with inline function.
[__sun]: Increase alignment to meet memalign precondition.
[!HAVE__ALIGNED_MALLOC && !HAVE_POSIX_MEMALIGN && !HAVE_MEMALIGN]
(aligned_alloc): Move check for valid alignment to operator new.
Remove redundant check for non-zero size, it's enforced by the caller.
(operator new): Move check for valid alignment here. Use
__builtin_expect on check for zero size.

From-SVN: r263360

Fix PR number for HPPA speculation patch: PR target/86807 ->  PR target/86785

This just fixes the PR number in the ChangeLog.  Nothing we can do
about the SVN history.

From-SVN: r263358

re PR c++/79133 (lambda capture shadowing parameter & decltype confusion)

PR c++/79133

gcc/cp/

PR c++/79133
* name-lookup.c (check_local_shadow): Reject captures and parameters
with the same name.

testsuite/

PR c++/79133
* g++.dg/cpp0x/lambda/lambda-shadow3.C: New.
* g++.dg/cpp1y/lambda-generic-variadic18.C: Likewise.

From-SVN: r263357

Add malloc predictor (PR middle-end/83023).

2018-08-07  Martin Liska  <mliska@suse.cz>

        PR middle-end/83023
* predict.c (expr_expected_value_1): Handle DECL_IS_MALLOC,
        BUILT_IN_REALLOC and DECL_IS_OPERATOR_NEW.
* predict.def (PRED_MALLOC_NONNULL): New predictor.
* doc/extend.texi: Document that malloc attribute adds
        hit to compiler.
2018-08-07  Martin Liska  <mliska@suse.cz>

        PR middle-end/83023
* gcc.dg/predict-16.c: New test.
* g++.dg/predict-1.C: New test.

From-SVN: r263355

Define monotonic_buffer_resource members out-of-line

Move the allocation logic into libstdc++.so so that it can be changed
without worrying about inlined code in existing binaries.

Leave do_allocate inline so that calls to it can be devirtualized, and
only the slow path needs to call into the library.

* config/abi/pre/gnu.ver: Export monotonic_buffer_resource members.
* include/std/memory_resource (monotonic_buffer_resource::release):
Call _M_release_buffers to free buffers.
(monotonic_buffer_resource::do_allocate): Call _M_new_buffer to
allocate a new buffer from upstream.
(monotonic_buffer_resource::_M_new_buffer): Declare.
(monotonic_buffer_resource::_M_release_buffers): Declare.
(monotonic_buffer_resource::_Chunk): Replace definition with
declaration as opaque type.
* src/c++17/memory_resource.cc (monotonic_buffer_resource::_Chunk):
Define.
(monotonic_buffer_resource::_M_new_buffer): Define.
(monotonic_buffer_resource::_M_release_buffers): Define.

From-SVN: r263354

Fix gcc.dg/vect/no-section-anchors-vect-69.c on SPARC etc. (PR tree-optimization/80925)

2018-08-07  Steve Ellcey  <sellcey@cavium.com>
    Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

PR tree-optimization/80925
* gcc.dg/vect/no-section-anchors-vect-69.c: Expect 3 loops
vectorized on !vect_hw_misalign targets.

Co-Authored-By: Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
From-SVN: r263352

Add missing gcc-interface/ to 2018-07-31 ChangeLog entry

From-SVN: r263351

Daily bump.

From-SVN: r263350

[libgomp, nvptx] Handle CUDA_ONE_CALL_MAYBE_NULL

This patch adds handling of functions that may not be present in the cuda
driver.

Such a function can be declared using CUDA_ONE_CALL_MAYBE_NULL in cuda-lib.def,
it can be called with the usual convenience macros, but before calling its
presence needs to be tested using new macro CUDA_CALL_EXISTS.

When using the dlopen interface (PLUGIN_NVPTX_DYNAMIC == 1), we allow
non-present functions by allowing dlsym to return NULL.  Otherwise
(PLUGIN_NVPTX_DYNAMIC == 0) we declare the non-present function to be weak.

Build and reg-tested libgomp on x86_64 with nvidia accelerator, with and without
--disable-cuda-driver, in combination with a trigger patch that adds a
non-existing function foo to cuda-lib.def:
...
CUDA_ONE_CALL_MAYBE_NULL (foo)
...
and declares it in plugin-nvptx.c:
...
CUresult foo (void);
...
and then uses it in nvptx_init after the init_cuda_lib call:
...
  if (CUDA_CALL_EXISTS (foo))
    CUDA_CALL (foo);
...

Also build and reg-tested on x86_64 with nvidia accelerator, with and without
--disable-cuda-driver, in combination with a trigger patch that replaces all
CUDA_ONE_CALLs in cuda-lib.def with CUDA_ONE_CALL_MAYBE_NULL, and guards two
CUDA_CALLs with CUDA_CALL_EXISTS, one for a regular fn, and one for a fn that is
a define in cuda/cuda.h.

2018-08-07  Tom de Vries  <tdevries@suse.de>

* plugin/plugin-nvptx.c (DO_PRAGMA): Define.
(struct cuda_lib_s): Add def/undef of CUDA_ONE_CALL_MAYBE_NULL.
(init_cuda_lib): Add new param to CUDA_ONE_CALL_1.  Add arg to
corresponding call in CUDA_ONE_CALL.  Add def/undef of
CUDA_ONE_CALL_MAYBE_NULL.
(CUDA_CALL_EXISTS): Define.

From-SVN: r263346

[libgomp, nvptx] Minimize lifetime of CUDA_ONE_CALL defines

This patch makes sure that the lifetimes of the CUDA_ONE_CALL macro (which is
defined twice in plugin-nvptx.c) are minimized, to make it obvious that the
definitions are used only in the lib-cuda.def include.

Build on x86_64 with nvptx accelerator and reg-tested libgomp.

2018-08-07  Tom de Vries  <tdevries@suse.de>

* plugin/plugin-nvptx.c (struct cuda_lib_s, init_cuda_lib): Put
CUDA_ONE_CALL defines right before the cuda-lib.def include, and the
corresponding undefs right after.

From-SVN: r263345

re PR target/86807 (spu port needs updating for CVE-2017-5753)

PR target/86807
* config/pa/pa.c (TARGET_HAVE_SPECULATION_SAFE_VALUE):
Define to speculation_safe_value_not_needed.

From-SVN: r263344

tree-ssa-dom.c (dom_opt_dom_walker::optimize_stmt): Pass down the vr_values instance to cprop_into_stmt.

* tree-ssa-dom.c (dom_opt_dom_walker::optimize_stmt): Pass down
the vr_values instance to cprop_into_stmt.
(cprop_into_stmt): Pass vr_values instance down to cprop_operand.
(cprop_operand): Also query EVRP to determine if OP is a constant.

From-SVN: r263342

[PATCH] Diagnostic included-from loop

https://gcc.gnu.org/ml/gcc-patches/2018-08/msg00416.html
* diagnostic.c (diagnostic_report_current_module): Reroll
included-at loop.  Translate text.

From-SVN: r263341

re PR c++/86767 (continue statements in constexpr functions causes unbounded looping)

PR c++/86767
* constexpr.c (cxx_eval_statement_list): Handle continue.

* g++.dg/cpp1y/constexpr-86767.C: New test.

From-SVN: r263340

Fix memory leak in selftest::test_expansion_to_rtl

"make selftest-valgrind" shows:

187 bytes in 1 blocks are definitely lost in loss record 567 of 669
    at 0x4A081D4: calloc (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
    by 0x1F08260: xcalloc (xmalloc.c:162)
    by 0xB24F32: init_emit() (emit-rtl.c:5843)
    by 0xC10080: prepare_function_start() (function.c:4803)
    by 0xC10254: init_function_start(tree_node*) (function.c:4877)
    by 0x1CDF92A: selftest::test_expansion_to_rtl() (function-tests.c:595)
    by 0x1CE007C: selftest::function_tests_c_tests() (function-tests.c:676)
    by 0x1E010E7: selftest::run_tests() (selftest-run-tests.c:98)
    by 0x1062D1E: toplev::run_self_tests() (toplev.c:2225)
    by 0x1062F40: toplev::main(int, char**) (toplev.c:2303)
    by 0x1E5B90A: main (main.c:39)

The allocation in question is:

  crtl->emit.regno_pointer_align
    = XCNEWVEC (unsigned char, crtl->emit.regno_pointer_align_length);

This patch fixes this leak (and makes the output of
"make selftest-valgrind" clean) by calling free_after_compilation at the
end of the selftest in question.

gcc/ChangeLog:
* function-tests.c (selftest::test_expansion_to_rtl): Call
free_after_compilation.

From-SVN: r263339

pr86763.C (dg-additional-options): Add -lrt for target *-*-linux-gnu.

* g++.dg/torture/pr86763.C (dg-additional-options): Add -lrt
for target *-*-linux-gnu.

From-SVN: r263338

Enable clobber high for tls descs on Aarch64

gcc/
* config/aarch64/aarch64.md: Add clobber highs to tls_desc.

gcc/testsuite/
* gcc.target/aarch64/sve/tls_preserve_1.c: New test.
* gcc.target/aarch64/sve/tls_preserve_2.c: New test.
* gcc.target/aarch64/sve/tls_preserve_3.c: New test.

From-SVN: r263337

S/390: Don't unroll memory blk op loops

gcc/ChangeLog:

2018-08-06  Andreas Krebbel  <krebbel@linux.ibm.com>

* config/s390/s390.c (s390_loop_unroll_adjust): Prevent small
loops with memory block operations from getting unrolled.

gcc/testsuite/ChangeLog:

2018-08-06  Andreas Krebbel  <krebbel@linux.ibm.com>

* gcc.target/s390/nomemloopunroll-1.c: New test.

From-SVN: r263336

[spu, commit] Define TARGET_HAVE_SPECULATION_SAFE_VALUE

The SPU processor is not affected by speculation, so this macro can
safely be defined as speculation_safe_value_not_needed.

gcc/ChangeLog:

PR target/86807
* config/spu/spu.c (TARGET_HAVE_SPECULATION_SAFE_VALUE):
Define to speculation_safe_value_not_needed.

From-SVN: r263335

reload1.c (forget_old_reloads_1): Adjust CLOBBER_HIGH assert.

* reload1.c (forget_old_reloads_1): Adjust CLOBBER_HIGH
assert.

From-SVN: r263333

re PR target/86662 (msp430-elf segfault with -flto and -mlarge)

PR target/86662

        * gcc/tree.c (build_common_tree_nodes): Initialize integer_types array
        with all enabled __intN types.

        * gcc/testsuite/gcc.target/msp430/pr86662.c: New test.

From-SVN: r263332

Remaining support for clobber high

gcc/
* alias.c (record_set): Check for clobber high.
* cfgexpand.c (expand_gimple_stmt): Likewise.
* combine-stack-adj.c (single_set_for_csa): Likewise.
* combine.c (find_single_use_1): Likewise.
(set_nonzero_bits_and_sign_copies): Likewise.
(get_combine_src_dest): Likewise.
(is_parallel_of_n_reg_sets): Likewise.
(try_combine): Likewise.
(record_dead_and_set_regs_1): Likewise.
(reg_dead_at_p_1): Likewise.
(reg_dead_at_p): Likewise.
* dce.c (deletable_insn_p): Likewise.
(mark_nonreg_stores_1): Likewise.
(mark_nonreg_stores_2): Likewise.
* df-scan.c (df_find_hard_reg_defs): Likewise.
(df_uses_record): Likewise.
(df_get_call_refs): Likewise.
* dwarf2out.c (mem_loc_descriptor): Likewise.
* haifa-sched.c (haifa_classify_rtx): Likewise.
* ira-build.c (create_insn_allocnos): Likewise.
* ira-costs.c (scan_one_insn): Likewise.
* ira.c (equiv_init_movable_p): Likewise.
(rtx_moveable_p): Likewise.
(interesting_dest_for_shprep): Likewise.
* jump.c (mark_jump_label_1): Likewise.
* postreload-gcse.c (record_opr_changes): Likewise.
* postreload.c (reload_cse_simplify): Likewise.
(struct reg_use): Add source expr.
(reload_combine): Check for clobber high.
(reload_combine_note_use): Likewise.
(reload_cse_move2add): Likewise.
(move2add_note_store): Likewise.
* print-rtl.c (print_pattern): Likewise.
* recog.c (decode_asm_operands): Likewise.
(store_data_bypass_p): Likewise.
(if_test_bypass_p): Likewise.
* regcprop.c (kill_clobbered_value): Likewise.
(kill_set_value): Likewise.
* reginfo.c (reg_scan_mark_refs): Likewise.
* reload1.c (maybe_fix_stack_asms): Likewise.
(eliminate_regs_1): Likewise.
(elimination_effects): Likewise.
(mark_not_eliminable): Likewise.
(scan_paradoxical_subregs): Likewise.
(forget_old_reloads_1): Likewise.
* reorg.c (find_end_label): Likewise.
(try_merge_delay_insns): Likewise.
(redundant_insn): Likewise.
(own_thread_p): Likewise.
(fill_simple_delay_slots): Likewise.
(fill_slots_from_thread): Likewise.
(dbr_schedule): Likewise.
* resource.c (update_live_status): Likewise.
(mark_referenced_resources): Likewise.
(mark_set_resources): Likewise.
* rtl.c (copy_rtx): Likewise.
* rtlanal.c (reg_referenced_p): Likewise.
(single_set_2): Likewise.
(noop_move_p): Likewise.
(note_stores): Likewise.
* sched-deps.c (sched_analyze_reg): Likewise.
(sched_analyze_insn): Likewise.

From-SVN: r263331

cse support for clobber_high

gcc/
* cse.c (invalidate_reg): New function extracted from...
(invalidate): ...here.
(canonicalize_insn): Check for clobber high.
(invalidate_from_clobbers): invalidate clobber highs.
(invalidate_from_sets_and_clobbers): Likewise.
(count_reg_usage): Check for clobber high.
(insn_live_p): Likewise.
* cselib.c (cselib_expand_value_rtx_1):Likewise.
(cselib_invalidate_regno): Check for clobber in setter.
(cselib_invalidate_rtx): Pass through setter.
(cselib_invalidate_rtx_note_stores):
(cselib_process_insn): Check for clobber high.
* cselib.h (cselib_invalidate_rtx): Add operand.

From-SVN: r263330

lra support for clobber_high

gcc/
* lra-eliminations.c (lra_eliminate_regs_1): Check for clobber high.
(mark_not_eliminable): Likewise.
* lra-int.h (struct lra_insn_reg): Add clobber high marker.
* lra-lives.c (process_bb_lives): Check for clobber high.
* lra.c (new_insn_reg): Remember clobber highs.
(collect_non_operand_hard_regs): Check for clobber high.
(lra_set_insn_recog_data): Likewise.
(add_regs_to_insn_regno_info): Likewise.
(lra_update_insn_regno_info): Likewise.

From-SVN: r263329

Add func to check if register is clobbered by clobber_high

gcc/
* rtl.h (reg_is_clobbered_by_clobber_high): Add declarations.
* rtlanal.c (reg_is_clobbered_by_clobber_high): Add function.

From-SVN: r263328

Generation support for CLOBBER_HIGH

Ensure clobber high is a register expression.
Info is passed through for the error case.

gcc/
* emit-rtl.c (verify_rtx_sharing): Check for CLOBBER_HIGH.
(copy_insn_1): Likewise.
(gen_hard_reg_clobber_high): New gen function.
* genconfig.c (walk_insn_part): Check for CLOBBER_HIGH.
* genemit.c (gen_exp): Likewise.
(gen_emit_seq): Pass through info.
(gen_insn): Check for CLOBBER_HIGH.
(gen_expand): Pass through info.
(gen_split): Likewise.
(output_add_clobbers): Likewise.
* genrecog.c (validate_pattern): Check for CLOBBER_HIGH.
(remove_clobbers): Likewise.
* rtl.h (gen_hard_reg_clobber_high): New declaration.

From-SVN: r263327

Add CLOBBER_HIGH expression

Includes documentation.

2018-08-06  Alan Hayward  <alan.hayward@arm.com>

* doc/rtl.texi (clobber_high): Add.
(parallel): Add in clobber high
* rtl.c (rtl_check_failed_code3): Add function.
* rtl.def (CLOBBER_HIGH): Add expression.
* rtl.h (RTL_CHECKC3): Add macro.
(rtl_check_failed_code3): Add declaration.
(XC3EXP): Add macro.

From-SVN: r263326

MAINTAINERS: Update my email address.

* MAINTAINERS: Update my email address.

From-SVN: r263325

Daily bump.

From-SVN: r263324

libbacktrace: Suppress the default action-if-found for AC_CHECK_LIBS.

Zlib is not a dependency of libbacktrace, and so it shouldn't be added
to LIBS.

libbacktrace/

* configure.ac: Move define of HAVE_ZLIB into check for -lz.
* Makefile.in: Regenerate.
* config.h.in: Likewise.
* configure: Likewise.

From-SVN: r263320

stl_iterator.h: Fix comment.

2018-08-05  François Dumont  <fdumont@gcc.gnu.org>

* include/bits/stl_iterator.h: Fix comment.

From-SVN: r263318

i386: Set cfun->machine->max_used_stack_alignment if needed

cfun->machine->max_used_stack_alignment is used to decide how stack frame
should be aligned.  This is independent of any psABIs nor 32-bit vs 64-bit.
It is always safe to compute max_used_stack_alignment.  We compute it only
if 128-bit aligned load/store may be generated on misaligned stack slot
which will lead to segfault.

gcc/

PR target/86386
* config/i386/i386.c (ix86_finalize_stack_frame_flags): Set
cfun->machine->max_used_stack_alignment if needed.

gcc/testsuite/

PR target/86386
* gcc.target/i386/pr86386.c: New file.

From-SVN: r263317

Daily bump.

From-SVN: r263316

PR tree-optimization/86571 - AIX NaNQ and NaNS output format conflicts with __builtin_sprintf

gcc/ChangeLog:

PR tree-optimization/86571
* gimple-ssa-sprintf.c (format_floating): Extend upper bound of
NaN output to 4.

From-SVN: r263312

[testsuite, guality] Use absolute line number in pass/fail line

2018-08-04  Tom de Vries  <tdevries@suse.de>

* lib/gcc-gdb-test.exp: Use absolute line number in pass/fail line.

From-SVN: r263311

[libgomp, nvptx, --without-cuda-driver] Don't use system cuda driver

Using libgomp configure option --with-cuda-driver=<dir> we can indicate what
cuda driver to use to build the libgomp nvptx plugin.  Without such an option,
the system cuda driver is used, if available.  If not availabe, a dlopen
interface is used instead.

However, when we use --without-cuda-driver (or the equivalent
--with-cuda-driver=no) the system cuda driver is still used if available.

This patch fixes that, making sure that --without-cuda-driver selects the dlopen
interface.

Build on x86_64 with nvptx accelerator and tested libgomp testsuite, with and
without option --without-cuda-driver.

2018-08-04  Tom de Vries  <tdevries@suse.de>

* plugin/configfrag.ac: For --without-cuda-driver, set
CUDA_DRIVER_INCLUDE and CUDA_DRIVER_LIB to no.  Handle
CUDA_DRIVER_INCLUDE == no and CUDA_DRIVER_LIB == no.
* configure: Regenerate.

From-SVN: r263310

const-volatile.c: Remove duplicate test "type:cvip".

* gcc.dg/guality/const-volatile.c: Remove duplicate test
"type:cvip".

From-SVN: r263309

re PR fortran/45521 ([F08] GENERIC resolution with ALLOCATABLE/POINTER and PROCEDURE)

2018-08-04  Janus Weil  <janus@gcc.gnu.org>

PR fortran/45521
* interface.c (gfc_compare_interfaces): Apply additional
distinguishability criteria of F08 to operator interfaces.

2018-08-04  Janus Weil  <janus@gcc.gnu.org>

PR fortran/45521
* gfortran.dg/interface_assignment_6.f90: New test case.

From-SVN: r263308

re PR testsuite/86153 (test case g++.dg/pr83239.C fails starting with r261585)

PR testsuite/86153
* g++.dg/pr83239.C (dg-options): Add -finline-limit=500.

From-SVN: r263306

Daily bump.

From-SVN: r263305

nios2.c (TARGET_HAVE_SPECULATION_SAFE_VALUE): Define.

2018-08-03  Sandra Loosemore  <sandra@codesourcery.com>

gcc/
* config/nios2/nios2.c (TARGET_HAVE_SPECULATION_SAFE_VALUE):
Define.

From-SVN: r263301

Remove nfs bogon

From-SVN: r263300

sjlj.S: Adjust to use PIC vs normal code to avoid absolute relocation in a shared library.

* config/sh/sjlj.S: Adjust to use PIC vs normal code to avoid
absolute relocation in a shared library.

From-SVN: r263299

Add fix-it hint for missing return statement in assignment operators (PR c++/85523)

gcc/cp/ChangeLog:

PR c++/85523
* decl.c: Include "gcc-rich-location.h".
(add_return_star_this_fixit): New function.
(finish_function): When warning about missing return statements in
functions returning non-void, add a "return *this;" fix-it hint for
assignment operators.

gcc/testsuite/ChangeLog:

PR c++/85523
* g++.dg/pr85523.C: New test.

Co-Authored-By: Jonathan Wakely <jwakely@redhat.com>
From-SVN: r263298

re PR target/86795 (mn10300 port needs updating for CVE-2017-5753)

PR target/86795
* config/mn10300/mn10300.c (TARGET_HAVE_SPECULATION_SAFE_VALUE):
Define to speculation_safe_value_not_needed.

From-SVN: r263296

docs: fix stray duplicated words

gcc/ChangeLog:
* doc/gcov.texi (-x): Remove duplicate "to".
* doc/invoke.texi (-Wnoexcept-type): Remove duplicate "calls".
(-Wif-not-aligned): Remove duplicate "is".
(-flto): Remove duplicate "the".
(MicroBlaze Options): In examples of "-mcpu=cpu-type", remove
duplicate "v5.00.b".
(MSP430 Options): Remove duplicate "and" from the description
of "-mgprel-sec=regexp".
(x86 Options): Remove duplicate copies of "vmldLog102" and
vmlsLog104 from description of "-mveclibabi=type".

From-SVN: r263295

Avoid infinite loop with duplicate anonymous union fields (PR c/86690).

If a struct contains an anonymous union and both have a field with the
same name, detect_field_duplicates_hash() will replace one of them
with NULL.  If compilation doesn't stop immediately, it may later call
lookup_field() on the union, which falsely assumes the union's
LANG_SPECIFIC array is sorted, and may loop indefinitely because of
this.

2018-08-03  Bogdan Harjoc  <harjoc@gmail.com>

PR c/86690
gcc/c:
* c-typeck.c (lookup_field): Do not use TYPE_LANG_SPECIFIC after
errors.

gcc/testsuite:
* gcc.dg/union-duplicate-field.c: New test.

From-SVN: r263294

re PR c++/86706 (ICE in build_base_path, at cp/class.c:294)

PR c++/86706
* class.c (build_base_path): Use currently_open_class.

* g++.dg/template/pr86706.C: New test.

From-SVN: r263293

ChangeLog: Move entry ...

* ChangeLog: Move entry ...
* cp/ChangeLog: ... here.

From-SVN: r263292

[Ada] Partially revert "Makefile patches from initial RISC-V cross/native build."

This partially reverts r262482, at it broke canadian builds.

2018-08-03  Pierre-Marie de Rodat  <derodat@adacore.com>

gcc/ada/

Reverts
2018-07-06  Jim Wilson  <jimw@sifive.com>

* Make-generated.in (treeprs.ads): Use $(GNATMAKE) instead of gnatmake.
(einfo.h, sinfo.h, stamp-snames, stamp-nmake): Likewise.
* gcc-interface/Makefile.in (xoscons): Likewise.

From-SVN: r263291

Handle SLP of call pattern statements

We couldn't vectorise:

  for (int j = 0; j < n; ++j)
    {
      for (int i = 0; i < 16; ++i)
a[i] = (b[i] + c[i]) >> 1;
      a += step;
      b += step;
      c += step;
    }

at -O3 because cunrolli unrolled the inner loop and SLP couldn't handle
AVG_FLOOR patterns (see also PR86504).  The problem was some overly
strict checking of pattern statements compared to normal statements
in vect_get_and_check_slp_defs:

          switch (gimple_code (def_stmt))
            {
            case GIMPLE_PHI:
            case GIMPLE_ASSIGN:
      break;

    default:
      if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
 "unsupported defining stmt:\n");
      return -1;
            }

The easy fix would have been to add GIMPLE_CALL to the switch,
but I don't think the switch is doing anything useful.  We only create
pattern statements that the rest of the vectoriser can handle, and the
other checks in this function and elsewhere check whether SLP is possible.

I'm also not sure why:

          if (!first && !oprnd_info->first_pattern
      /* Allow different pattern state for the defs of the
 first stmt in reduction chains.  */
      && (oprnd_info->first_dt != vect_reduction_def

is necessary.  All that should matter is that the statements in the
node are "similar enough".  It turned out to be quite hard to find a
convincing example that used a mixture of pattern and non-pattern
statements, so bb-slp-pow-1.c is the best I could come up with.
But it does show that the combination of "xi * xi" statements and
"pow (xj, 2) -> xj * xj" patterns are handled correctly.

The patch therefore just removes the whole if block.

The loop also needed commutative swapping to be extended to at least
AVG_FLOOR.

This gives +3.9% on 525.x264_r at -O3.

2018-08-03  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* internal-fn.h (first_commutative_argument): Declare.
* internal-fn.c (first_commutative_argument): New function.
* tree-vect-slp.c (vect_get_and_check_slp_defs): Remove extra
restrictions for pattern statements.  Use first_commutative_argument
to look for commutative operands in calls to internal functions.

gcc/testsuite/
* gcc.dg/vect/bb-slp-over-widen-1.c: Expect AVG_FLOOR to be used
on vect_avg_qi targets.
* gcc.dg/vect/bb-slp-over-widen-2.c: Likewise.
* gcc.dg/vect/bb-slp-pow-1.c: New test.
* gcc.dg/vect/vect-avg-15.c: Likewise.

From-SVN: r263290

Add workaround for non-unique errno values on AIX

* src/c++11/system_error.cc
(system_error_category::default_error_condition): Add workaround for
ENOTEMPTY and EEXIST having the same value on AIX.
* testsuite/19_diagnostics/error_category/system_category.cc: Add
extra testcases for EDOM, EILSEQ, ERANGE, EEXIST and ENOTEMPTY.

From-SVN: r263289

Makefile.in (wide-int-range.o): New.

* Makefile.in (wide-int-range.o): New.
* tree-vrp.c: Move all the wide_int_* functions to...
* wide-int-range.cc: ...here.
* tree-vrp.h: Move all the wide_int_* prototypes to...
* wide-int-range.h: ...here.

From-SVN: r263288

[c++] Don't emit exception tables for UI_NONE

If a target does not support exceptions, it can indicate this by returning
UI_NONE in TARGET_EXCEPT_UNWIND_INFO.  Currently the compiler still emits
exception tables for such a target.

This patch makes sure that no exception tables are emitted if the target does
not support exceptions.  This allows us to remove a workaround in
TARGET_ASM_BYTE_OP in the nvptx port.

Build on x86_64 with nvptx accelerator, and tested libgomp.

Build and reg-tested on x86_64.

2018-08-03  Tom de Vries  <tdevries@suse.de>

* common/config/nvptx/nvptx-common.c (nvptx_except_unwind_info): Return
UI_NONE.
* config/nvptx/nvptx.c (TARGET_ASM_BYTE_OP): Remove define.
* except.c (output_function_exception_table): Do early exit if
targetm_common.except_unwind_info (&global_options) == UI_NONE.

From-SVN: r263287

Print heuristics probability fraction part with 2 digits.

2018-08-03  Martin Liska  <mliska@suse.cz>

* predict.c (dump_prediction): Change to 2 digits
        in fraction part.
2018-08-03  Martin Liska  <mliska@suse.cz>

* gcc.dg/predict-1.c: Adjust scanned pattern to cover 2 digits.
* gcc.dg/predict-13.c:Likewise.
* gcc.dg/predict-3.c:Likewise.
* gcc.dg/predict-4.c:Likewise.
* gcc.dg/predict-5.c:Likewise.
* gcc.dg/predict-6.c:Likewise.
* gcc.dg/predict-9.c:Likewise.
* gfortran.dg/predict-1.f90:Likewise.

From-SVN: r263286

[aarch64] Fix falkor pipeline description for dup<q>

There was a typo in the pipeline description where DUP was assigned to
the vector pipes for quad mode ops when it really only uses the VTOG
pipes.  Fixing this does not show any noticeable difference in
performance (there's a very small bump of 1.7% in x264 but that's
about it) in my tests but is the more precise description of operations
for falkor.

* config/aarch64/falkor.md (falkor_am_1_vxvy_vxvy): Move
neon_dup_q to...
(falkor_am_1_gtov_gtov): ... a new insn reservation.

From-SVN: r263285

Daily bump.

From-SVN: r263284

nds32.c (nds32_hard_regno_mode_ok): Replace > with >=.

* config/nds32/nds32.c (nds32_hard_regno_mode_ok): Replace > with >=.
* df-problems.c (df_remove_dead_eq_notes): Replace > with >=.
* dwarf2out.c (mem_loc_descriptor): Replace > with >=.
* lra-constraints.c (spill_hard_reg_in_range): Replace <= with <.
* lra-remat.c (call_used_input_regno_present_p): Replace <= with <.

From-SVN: r263280

Fix memory leak of pretty_printer prefixes

We were rather sloppy about handling the ownership of prefixes for
pretty_printer, and this lead to a memory leak for any time a
diagnostic_show_locus call emits multiple line spans.

This showed up in "make selftest-valgrind" as:

3,976 bytes in 28 blocks are definitely lost in loss record 632 of 669
   at 0x4A0645D: malloc (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
   by 0x1F08227: xmalloc (xmalloc.c:147)
   by 0x1F083E6: xvasprintf (xvasprintf.c:58)
   by 0x1E7EC7D: build_message_string(char const*, ...) (diagnostic.c:78)
   by 0x1E7F438: diagnostic_get_location_text(diagnostic_context*, expanded_location) (diagnostic.c:328)
   by 0x1E7FD54: default_diagnostic_start_span_fn(diagnostic_context*, expanded_location) (diagnostic.c:626)
   by 0x1EB3508: selftest::test_diagnostic_context::start_span_cb(diagnostic_context*, expanded_location) (selftest-diagnostic.c:57)
   by 0x1E89215: diagnostic_show_locus(diagnostic_context*, rich_location*, diagnostic_t) (diagnostic-show-locus.c:1992)
   by 0x1E8ECAD: selftest::test_fixit_insert_containing_newline_2(selftest::line_table_case const&) (diagnostic-show-locus.c:3044)
   by 0x1EB0606: selftest::for_each_line_table_case(void (*)(selftest::line_table_case const&)) (input.c:3525)
   by 0x1E8F3F5: selftest::diagnostic_show_locus_c_tests() (diagnostic-show-locus.c:3164)
   by 0x1E010BF: selftest::run_tests() (selftest-run-tests.c:88)

4,004 bytes in 28 blocks are definitely lost in loss record 633 of 669
   at 0x4A0645D: malloc (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
   by 0x1F08227: xmalloc (xmalloc.c:147)
   by 0x1F083E6: xvasprintf (xvasprintf.c:58)
   by 0x1E7EC7D: build_message_string(char const*, ...) (diagnostic.c:78)
   by 0x1E7F438: diagnostic_get_location_text(diagnostic_context*, expanded_location) (diagnostic.c:328)
   by 0x1E7FD54: default_diagnostic_start_span_fn(diagnostic_context*, expanded_location) (diagnostic.c:626)
   by 0x1EB3508: selftest::test_diagnostic_context::start_span_cb(diagnostic_context*, expanded_location) (selftest-diagnostic.c:57)
   by 0x1E89215: diagnostic_show_locus(diagnostic_context*, rich_location*, diagnostic_t) (diagnostic-show-locus.c:1992)
   by 0x1E8B373: selftest::test_diagnostic_show_locus_fixit_lines(selftest::line_table_case const&) (diagnostic-show-locus.c:2500)
   by 0x1EB0606: selftest::for_each_line_table_case(void (*)(selftest::line_table_case const&)) (input.c:3525)
   by 0x1E8F3B9: selftest::diagnostic_show_locus_c_tests() (diagnostic-show-locus.c:3159)
   by 0x1E010BF: selftest::run_tests() (selftest-run-tests.c:88)

This patch fixes the leaks by ensuring that the pretty_printer "owns"
the prefix if it's non-NULL, freeing it in the dtor and in pp_set_prefix.

gcc/cp/ChangeLog:
* error.c (cxx_print_error_function): Duplicate "file" before
passing it to pp_set_prefix.
(cp_print_error_function): Use pp_take_prefix when saving the
existing prefix.

gcc/ChangeLog:
* diagnostic-show-locus.c (diagnostic_show_locus): Use
pp_take_prefix when saving the existing prefix.
* diagnostic.c (diagnostic_append_note): Likewise.
* langhooks.c (lhd_print_error_function): Likewise.
* pretty-print.c (pp_set_prefix): Drop the "const" from "prefix"
param's type.  Free the existing prefix.
(pp_take_prefix): New function.
(pretty_printer::pretty_printer): Drop the prefix parameter.
Rename the length parameter to match the comment.
(pretty_printer::~pretty_printer): Free the prefix.
* pretty-print.h (pretty_printer::pretty_printer): Drop the prefix
parameter.
(struct pretty_printer): Drop the "const" from "prefix" field's
type and clarify memory management.
(pp_set_prefix): Drop the "const" from the 2nd param.
(pp_take_prefix): New decl.

From-SVN: r263275