PR c++/90875 - added -Wswitch-outside-range option

From-SVN: r272567

[Darwin, fixincludes] Fix Darwin9/10 math.h issues.

Darwin has had long long functions for some considerable time and these are
exposed in Darwin8 and Darwin11+ headers.  However, for some reason it was
elected to hide them behind __STRICT_ANSI__ and __STDC_VERSION__ on Darwin9
and Darwin10.  This is a problem for G++/libstdc++ that expects the functions
to be available for strict ansi (-std=c++14, for example) and without
defining __STDC_VERSION__.  The fix here follows the pattern used in
Darwin11+ headers where the functions may be explicitly hidden by defining
__DARWIN_NO_LONG_LONG.

This fixes the tr1 testsuite fails seen on Darwin9 and 10.

2019-06-21  Iain Sandoe  <iain@sandoe.co.uk>

* inclhack.def: Replace the complex test using __STRICT_ANSI__ and
__STDC_VERSION__ with a test using __DARWIN_NO_LONG_LONG.
Ensure that the top level math.h uses <> to wrap included headers
rather than "".
* fixincl.x: Regenerated.
* tests/base/architecture/ppc/math.h: Update test to include the
__DARWIN_NO_LONG_LONG case.

From-SVN: r272563

[Darwin, fixincludes] Fix Darwin14 header issues.

There are two issues with the Darwin14 (SDK) headers in which unguarded
advanced syntax elements causes any code including these headers to fail.

2019-06-21  Iain Sandoe  <iain@sandoe.co.uk>

* inclhack.def: Guard __has_attribute and __has_extension in
os/base.h.
Guard Apple blocks syntax in dispatch/object.h.
* fixincl.x: Regenerate.
* tests/base/dispatch/object.h: New file.
* tests/base/os/base.h: New file.

From-SVN: r272561

re PR tree-optimization/90930 (Excessive memory consumption)

2019-06-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/90930
* tree-ssa-reassoc.c (rewrite_expr_tree_parallel): Set visited
flag on new stmts to avoid re-processing them.

From-SVN: r272560

PR c++/90875 - added -Wswitch-outside-range option

* doc/invoke.texi (Wswitch-outside-range): Document.

* c.opt (Wswitch-outside-range): Added new option.
* c-warn.c (c_do_switch_warnings): Use OPT_Wswitch-outside-range.

* c-c++-common/Wswitch-outside-range-1.c: New test.
* c-c++-common/Wswitch-outside-range-2.c: New test.
* c-c++-common/Wswitch-outside-range-3.c: New test.
* c-c++-common/Wswitch-outside-range-4.c: New test.

From-SVN: r272559

re PR fortran/51991 (Wrong error message with variables named "SAVE*")

2019-06-21  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/51991
* decl.c (gfc_match_save): If SAVE was not seen, return MATCH_NO
instead issuing an error message and returning MATCH_ERROR.

2019-06-21  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/51991
gfortran.dg/pr51991.f90

From-SVN: r272556

re PR tree-optimization/90949 (null pointer check removed)

PR tree-optimization/90949
* tree-ssa-copy.c (fini_copy_prop): Use reset_flow_sensitive_info.
* tree-ssanames.c (reset_flow_sensitive_info): Reset non-null state.

* gcc.c-torture/execute/pr90949.c: New test.

From-SVN: r272555

Fix missing name in ChangeLog

Fix another case in which a name in the commit message wasn't copied to
the ChangeLog -- sorry!

From-SVN: r272553

PR c++/90953 - ICE with -Wmissing-format-attribute.

* c-common.c (check_function_arguments_recurse): Use
get_attribute_name.
(check_missing_format_attribute): Likewise.

* g++.dg/warn/miss-format-7.C: New test.

From-SVN: r272552

Fix missing name in ChangeLog

It was in the commit message but got dropped from the ChangeLog entry
due to a scripting bug.

From-SVN: r272551

compiler: omit write barrier for assignment to *(convert(&local))

    Assignments to local variables don't need a write barrier. But
    currently the compiler inserts a write barrier if the LHS is a
    local variable with type converted, as *(convert(&local)). Let
    the compiler recognize this pattern and omit the write barrier.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/182541

From-SVN: r272550

compiler: open code string slice expressions

    Currently a string slice expression is implemented with a runtime
    call __go_string_slice. Change it to open code it, which is more
    efficient, and allows the backend to further optimize it.

    Also omit the write barrier for length-only update (i.e.
    s = s[:n]).

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/182540

From-SVN: r272549

re PR debug/90914 (ICE in schedule_generic_params_dies_gen, at dwarf2out.c:27153)

2019-06-21  Richard Biener  <rguenther@suse.de>

PR debug/90914
* dwarf2out.c (prune_unused_types_walk): Always consider
function-local extern declarations as used.

* g++.dg/debug/pr90914.C: New testcase.

From-SVN: r272547

re PR tree-optimization/90913 (ICE in maybe_gen_insn, at optabs.c:7341 since r272239)

2019-06-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/90913
* tree-vect-loop-manip.c (vect_loop_versioning): Do not re-use
the scalar variant of if-conversion versioning.

* gfortran.dg/vect/pr90913.f90: New testcase.

From-SVN: r272545

omp-low.c (lower_rec_simd_input_clauses): Add rvar2 argument...

* omp-low.c (lower_rec_simd_input_clauses): Add rvar2 argument,
create another "omp scan inscan exclusive" array if
!ctx->scan_inclusive.
(lower_rec_input_clauses): Handle exclusive scan inscan reductions.
(lower_omp_scan): Likewise.
* tree-vectorizer.h (struct _stmt_vec_info): Use 3-bit instead of
2-bit bitfield for simd_lane_access_p member.
* tree-vect-data-refs.c (vect_analyze_data_refs): Also handle
aux == (void *)-4 as simd lane access.
* tree-vect-stmts.c (check_scan_store): Handle exclusive scan.  Update
comment with permutations to show the canonical permutation order.
(vectorizable_scan_store): Handle exclusive scan.
(vectorizable_store): Call vectorizable_scan_store even for
STMT_VINFO_SIMD_LANE_ACCESS_P > 3.

* gcc.dg/vect/vect-simd-12.c: New test.
* gcc.dg/vect/vect-simd-13.c: New test.
* gcc.dg/vect/vect-simd-14.c: New test.
* gcc.dg/vect/vect-simd-15.c: New test.
* gcc.target/i386/sse2-vect-simd-12.c: New test.
* gcc.target/i386/sse2-vect-simd-13.c: New test.
* gcc.target/i386/sse2-vect-simd-14.c: New test.
* gcc.target/i386/sse2-vect-simd-15.c: New test.
* gcc.target/i386/avx2-vect-simd-12.c: New test.
* gcc.target/i386/avx2-vect-simd-13.c: New test.
* gcc.target/i386/avx2-vect-simd-14.c: New test.
* gcc.target/i386/avx2-vect-simd-15.c: New test.
* gcc.target/i386/avx512f-vect-simd-12.c: New test.
* gcc.target/i386/avx512f-vect-simd-13.c: New test.
* gcc.target/i386/avx512f-vect-simd-14.c: New test.
* gcc.target/i386/avx512bw-vect-simd-15.c: New test.
* g++.dg/vect/simd-6.cc: New test.
* g++.dg/vect/simd-7.cc: New test.
* g++.dg/vect/simd-8.cc: New test.
* g++.dg/vect/simd-9.cc: New test.
* c-c++-common/gomp/scan-2.c: Don't expect any diagnostics.

From-SVN: r272544

re PR c++/90950 (OpenMP clause handling rejecting references to incomplete types in templates)

PR c++/90950
* semantics.c (finish_omp_clauses): Don't reject references to
incomplete types if processing_template_decl.

* g++.dg/gomp/lastprivate-1.C: New test.

From-SVN: r272543

tree-vect-data-refs.c (vect_find_stmt_data_reference): Handle "omp simd array" arrays with one byte elements.

* tree-vect-data-refs.c (vect_find_stmt_data_reference): Handle
"omp simd array" arrays with one byte elements.

* gcc.dg/vect/vect-simd-11.c: New test.
* gcc.target/i386/sse2-vect-simd-11.c: New test.
* gcc.target/i386/avx2-vect-simd-11.c: New test.
* gcc.target/i386/avx512bw-vect-simd-11.c: New test.

From-SVN: r272542

Daily bump.

From-SVN: r272538

re PR c++/79781 (ICE on valid C++ code with -std=c++14 (in assemble_integer, at varasm.c:2733))

PR c++/79781
* g++.dg/ext/goto1.C: New test.

From-SVN: r272527

re PR fortran/77632 ([F08] Pointer initialisation does not quite work with arrays)

2019-06-20  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/77632
* /decl.c (variable_decl): Mark a variable that is a target in pointer
initialization when in PROGRAM, MODULE, or SUBMODULE scope with an
implicit save.

2019-06-20  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/77632
* gfortran.dg/pr77632_1.f90: New test.

From-SVN: r272526

re PR c++/68265 (Arbitrary syntactic nonsense silently accepted after 'int (*){}' until the next close brace)

PR c++/68265
* g++.dg/parse/error62.C: New test.

From-SVN: r272525

re PR fortran/86587 (Derived-type with attributes BIND(C) and PRIVATE raises an error but standard accepts it)

2019-06-20  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/86587
* symbol.c (verify_bind_c_derived_type): Remove erroneous error
checking for BIND(C) and PRIVATE attributes.

2019-06-20  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/86587
* gfortran.dg/pr86587.f90: New test.

From-SVN: r272524

alpha.md (@unaligned_store<mode>): Rename from unaligned_store<mode>.

* config/alpha/alpha.md (@unaligned_store<mode>):
Rename from unaligned_store<mode>.
(@reload_in<mode>_aligned): Rename from reload_in<mode>_aligned.
* config/alpha/sync.md (@load_locked_<mode>): Rename
from load_locked_<mode>.
(@store_conditional_<mode>): Rename from store_conditional_<mode>.
(@atomic_compare_and_swap<mode>_1): Rename
from atomic_compare_and_swap<mode>_1.
(@atomic_exchange<mode>_1): Rename from atomic_exchange<mode>_1.
* config/alpha/alpha.c (alpha_expand_mov_nobwx):
Use gen_reload_in_aligned and gen_unaligned_store.
(emit_load_locked): Remove.
(emit_store_conditional): Ditto.
(alpha_split_atomic_op): Use gen_load_locked and gen_store_conditional.
(alpha_split_compare_and_swap): Ditto.
(alpha_expand_compare_and_swap_12): Use gen_atomic_compare_and_swap_1.
(alpha_split_compare_and_swap_12): Use gen_load_locked
and gen_store_conditional.
(alpha_split_atomic_exchange): Ditto.
(alpha_expand_atomic_exchange_12): Use gen_atomic_exchange_1.
(alpha_split_atomic_exchange_12): Use gen_load_locked
and gen_store_conditional.

From-SVN: r272523

libfortran/65921: Add forgotten PR number to ChangeLog

2019-06-14  Janne Blomqvist  <jb@gcc.gnu.org>

        PR fortran/65921
        * runtime/memory.c (SIZE_MAX):Remove macro definition.
        (xmallocarray): Use __builtin_mul_overflow.

From-SVN: r272520

[objective-c/c++, testsuite] Fix stubify tests for -fnext-runtime.

These tests were broken in the first place, and more broken following
the change to drop pic symbol stubs unless the linker needs them.

Fixed to work for x86 as well as powerpc and amended the options to
force pic symbol stubs on.

gcc/testsuite/

2019-06-20  Iain Sandoe  <iain@sandoe.co.uk>

* obj-c++.dg/stubify-1.mm: Adjust options and scan-asm checks.
* obj-c++.dg/stubify-2.mm: Likewise.
* objc.dg/stubify-1.m: Likewise.
* objc.dg/stubify-2.m: Likewise.

From-SVN: r272518

[aarch64] Refactor common errata work-around specs

With the addition of the AArch64 NetBSD support code we now have four
ports all using and defining the same errata work-around headers.
That's silly and long-term becomes a maintenance burden.

This patch factors all that code into a single header to eliminate all
the duplication.

* config/aarch64/aarch64-errata.h: New file.
* config/aarch64/aarch64-elf-raw.h (CA53_ERR_835769_SPEC): Delete.
(CA53_ERR_843419_SPEC): Delete.
(LINK_SPEC): Use AARCH64_ERRATA_LINK_SPEC instead of above.
* config/aarch64/aarch64-linux.h: Likewise.
* config/aarch64/aarch64-netbsd.h: Likewise.
* config/aarch64/aarch64-freebsd.h: Likewise.

From-SVN: r272515

* config/sh/sh.c (sh2a_function_vector_p): Use get_attribute_name.

From-SVN: r272514

rs6000.md (isa attribute): Add support for for a future processor.

2019-06-20  Michael Meissner  <meissner@linux.ibm.com>

* config/rs6000/rs6000.md (isa attribute): Add support for
for a future processor.

From-SVN: r272513

re PR c++/87512 (Error: the type ‘const auto’ of ‘constexpr’ variable is not literal)

PR c++/87512
* g++.dg/cpp1z/inline-var7.C: New test.

From-SVN: r272512

i386: Generate standard floating point scalar operation patterns

Standard floating point scalar operation patterns for combiner, which
preserve the rest of the vector, look like

     (vec_merge:V2DF
       (vec_duplicate:V2DF (reg:DF 87))
       (reg/v:V2DF 85 [ x ])
       (const_int 1 [0x1])]))

and

     (vec_merge:V2DF
       (vec_duplicate:V2DF
         (op:DF (vec_select:DF (reg/v:V2DF 85 [ x ])
                (parallel [ (const_int 0 [0])]))
         (reg:DF 87))
       (reg/v:V2DF 85 [ x ])
       (const_int 1 [0x1])]))

This patch adds and generates such standard floating point scalar
operation patterns for +, -, *, /, > and <.

Tested on x86-64.

gcc/

PR target/54855
* config/i386/i386-expand.c (ix86_expand_vector_set): Generate
standard scalar operation pattern for V2DF.
* config/i386/sse.md (*<sse>_vm<plusminus_insn><mode>3): New.
(*<sse>_vm<multdiv_mnemonic><mode>3): Likewise.
(*ieee_<ieee_maxmin><mode>3): Likewise.
(vec_setv2df_0): Likewise.

gcc/testsuite/

PR target/54855
* gcc.target/i386/pr54855-1.c: New test.
* gcc.target/i386/pr54855-2.c: Likewise.
* gcc.target/i386/pr54855-3.c: Likewise.
* gcc.target/i386/pr54855-4.c: Likewise.
* gcc.target/i386/pr54855-5.c: Likewise.
* gcc.target/i386/pr54855-6.c: Likewise.
* gcc.target/i386/pr54855-7.c: Likewise.
* gcc.target/i386/pr54855-8.c: Likewise.
* gcc.target/i386/pr54855-9.c: Likewise.
* gcc.target/i386/pr54855-10.c: Likewise.

From-SVN: r272511

tree-ssa-alias.c (aliasing_component_refs_p): Remove ref2_is_decl parameter; it has no use in gimple memory model.

* tree-ssa-alias.c (aliasing_component_refs_p): Remove ref2_is_decl
parameter; it has no use in gimple memory model.
(indirect_ref_may_alias_decl_p): Update.

* gcc.c-torture/execute/alias-access-path-1.c: New testcase.

From-SVN: r272510

Skip libstdc++ debug build in early bootstrap stages

As mentioned in PR 90770, this is a patch that Debian have been carrying
for some time. The additional unoptimized copies of libstdc++ libs that
get built during each stage are never going to be used, so don't bother
building them.

For a profiled bootstrap this means we won't train the compiler on the
unoptimized library code with assertions enabled, but that doesn't seem
like a big problem, as the same code has already been compiled once for
the main libstdc++ library.

* acinclude.m4 (GLIBCXX_ENABLE_DEBUG): Only do debug build for final
stage of bootstrap.
* configure: Regenerate.

From-SVN: r272509

Qualify calls to __never_valueless in <variant>

* include/std/variant (_Variant_storage, _Extra_visit_slot_needed):
Qualify calls to __never_valueless.

From-SVN: r272508

re PR c++/89873 (internal compiler error: unexpected expression of kind implicit_conv_expr)

PR c++/89873
* g++.dg/cpp1y/noexcept1.C: New test.

From-SVN: r272507

re PR fortran/90937 (ICE: in gfc_get_symbol_decl, at fortran/trans-decl.c:1538)

2019-06-20  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/90937
* trans-types.c (get_formal_from_actual_arglist): Get symbol from
current namespace so it will be freed later.  If symbol is of type
character, get an empty character length.

2019-06-20  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/90937
* gfortran.dg/external_procedure_4.f90: New test.

From-SVN: r272506

[testsuite] Add missing dg-require-effective-target global_constructor

Add missing dg-require-effective-target global_constructor.

Tested on nvptx.

2019-06-20  Tom de Vries  <tdevries@suse.de>

* gcc.dg/pr90866-2.c: Require global_constructor.

From-SVN: r272505

[testsuite] Add missing dg-require-effective-target nonlocal_goto

Add missing dg-require-effective-target nonlocal_goto.

Tested on nvptx.

2019-06-20  Tom de Vries  <tdevries@suse.de>

* gcc.c-torture/compile/pr89280.c: Require nonlocal_goto.
* gcc.dg/pr88870.c: Same.
* gcc.dg/pr90082.c: Same.

From-SVN: r272504

[testsuite] Add missing dg-require-effective-target indirect_jumps

Add missing dg-require-effective-target indirect_jumps.

Tested on nvptx.

2019-06-20  Tom de Vries  <tdevries@suse.de>

* gcc.dg/pr89737.c: Require indirect_jumps.
* gcc.dg/torture/pr87693.c: Same.
* gcc.dg/torture/pr89135.c: Same.
* gcc.dg/torture/pr90071.c: Same.

From-SVN: r272503

[testsuite] Add missing dg-require-effective-target label_values

Add missing dg-require-effective-target label_values.

Tested on nvptx.

2019-06-20  Tom de Vries  <tdevries@suse.de>

* gcc.c-torture/compile/pr89280.c: Require label_values.
* gcc.dg/pr89737.c: Same.
* gcc.dg/pr90082.c: Same.
* gcc.dg/torture/pr89135.c: Same.
* gcc.dg/torture/pr89247.c: Same.
* gcc.dg/torture/pr90071.c: Same.

From-SVN: r272502

[testsuite] Add missing dg-require-effective-target alloca

Add missing dg-require-effective-target alloca.

Tested on nvptx.

2019-06-20  Tom de Vries  <tdevries@suse.de>

* gcc.c-torture/compile/pr77754-1.c: Require alloca.
* gcc.c-torture/compile/pr77754-2.c: Same.
* gcc.c-torture/compile/pr77754-3.c: Same.
* gcc.c-torture/compile/pr77754-4.c: Same.
* gcc.c-torture/compile/pr77754-5.c: Same.
* gcc.c-torture/compile/pr77754-6.c: Same.
* gcc.c-torture/compile/pr87110.c: Same.
* gcc.c-torture/execute/pr86528.c: Same.
* gcc.dg/Walloca-larger-than-2.c: Same.
* gcc.dg/Walloca-larger-than.c: Same.
* gcc.dg/Warray-bounds-41.c: Same.
* gcc.dg/Wrestrict-17.c: Same.
* gcc.dg/Wstrict-overflow-27.c: Same.
* gcc.dg/Wstringop-truncation-3.c: Same.
* gcc.dg/pr78902.c: Same.
* gcc.dg/pr87099.c: Same.
* gcc.dg/pr87320.c: Same.
* gcc.dg/pr89045.c: Same.
* gcc.dg/strlenopt-62.c: Same.
* gcc.dg/tree-ssa/alias-37.c: Same.

From-SVN: r272501

Fix outdated reference to C++17 draft in the docs

* doc/xml/manual/status_cxx2017.xml: Fix outdated reference to
C++17 working draft.

From-SVN: r272500

Improve tests for std::vector<bool> printer

The current tests wouldn't notice if the vector<bool> contents were
printed in reverse, because it would read the same forwards and
backwards. Change the content so the tests would fail if that happened.

* testsuite/libstdc++-prettyprinters/simple.cc: Use non-palindromic
vector<bool> for test.
* testsuite/libstdc++-prettyprinters/simple11.cc: Likewise.

From-SVN: r272499

Decrease hash-table-verification-limit from 100 to 10.

2019-06-20  Martin Liska  <mliska@suse.cz>

* params.def (PARAM_HASH_TABLE_VERIFICATION_LIMIT): Decrease
to 10.

From-SVN: r272498

tree-vect-stmts.c (enum scan_store_kind): New type.

* tree-vect-stmts.c (enum scan_store_kind): New type.
(scan_store_can_perm_p): Change last argument from int * to
vec<enum scan_store_kind> *, record precisely which permutations
need whole vector left shift or that plus VEC_COND_EXPR.
(vectorizable_scan_store): Adjust caller, use whole vector left shift
and additional VEC_COND_EXPR only for those iterations that need it.

From-SVN: r272497

fix ARM --with-fpu option checking and error message

Fix the test for failure in parsecpu's checking of the --with-fpu
argument, and the error message that gets printed when the check
fails.

for  gcc/ChangeLog

* config.gcc: Fix ARM --with-fpu checking and error message.

From-SVN: r272496

Daily bump.

From-SVN: r272495

Fix non-standard behaviour of std::istream_iterator

The current implementation of istream_iterator allows the iterator to be
reused after reaching end-of-stream, so that subsequent reads from the
stream can succeed (e.g. if the stream state has been cleared and stream
position changed from EOF). The P0738R2 paper clarified that the
expected behaviour is to set the stream pointer to null after reaching
end-of-stream, preventing further reads.

This implements that requirement, and adds the new default constructor
to std::ostream_iterator.

* include/bits/stream_iterator.h (istream_iterator::_M_equal()): Make
private.
(istream_iterator::_M_read()): Do not check stream state before
attempting extraction. Set stream pointer to null when extraction
fails (P0738R2).
(operator==(const istream_iterator&, const istream_iterator&)): Change
to be a hidden friend of istream_iterator.
(operator!=(const istream_iterator&, const istream_iterator&)):
Likewise.
(ostream_iterator::ostream_iterator()): Add default constructor.
(ostream_iterator::ostream_iterator(ostream_type*, const C*)): Use
addressof.
* testsuite/24_iterators/istream_iterator/1.cc: New test.
* testsuite/24_iterators/ostream_iterator/1.cc: New test.
* testsuite/24_iterators/ostream_iterator/70766.cc: Also check
constructor taking a string.
* testsuite/24_iterators/ostream_iterator/requirements/constexpr.cc:
New test.

From-SVN: r272491

Have std::vector printer's iterator return bool for vector<bool>

Have the pretty-printer for 'std::vector<bool>' return a
value of type 'bool' rather than an 'int'.

This way, the type is clear and that can be used for better
display and a 'gdb.Value' constructed from the returned value
will have type 'bool' again, not e.g. 'long long' as happened
previously (at least with GDB 8.2.1 on amd64).

2019-06-19  Michael Weghorn  <m.weghorn@posteo.de>
    Jonathan Wakely  <jwakely@redhat.com>

PR libstdc++/90945
* python/libstdcxx/v6/printers.py (StdVectorPrinter._iterator): Use
values of type bool for vector<bool> elements.
* testsuite/libstdc++-prettyprinters/simple.cc: Test vector<bool>.
* testsuite/libstdc++-prettyprinters/simple11.cc: Likewise.

Co-Authored-By: Jonathan Wakely <jwakely@redhat.com>
From-SVN: r272490

PR libstdc++/90920 restore previous checks for empty ranges

The change in r263433 broke the contract of the __rotate functions, by no
longer accepting empty ranges. That means that callers which inlined the
old version of std::rotate (without checks) that end up linking to a new
definition of std::__rotate (also without checks) could perform a divide
by zero and crash.

This restores the old contract of the __rotate overloads.

PR libstdc++/90920 partially revert r263433
* include/bits/stl_algo.h (__rotate): Restore checks for empty ranges.
(rotate): Remove checks.
* testsuite/25_algorithms/rotate/90920.cc: New test.

From-SVN: r272489

re PR tree-optimization/90626 (fold strcmp(a, b) == 0 to zero when one string length is exact and the other is unequal)

PR tree-optimization/90626

gcc/ChangeLog:
* tree-ssa-strlen.c (strxcmp_unequal): Fix typos.

From-SVN: r272487

PR c++/60364 - noreturn after first decl not diagnosed.

* attribs.c (get_attribute_namespace): No longer static.
(decl_attributes): Avoid shadowing.  Preserve the C++11 form for C++11
attributes.
(attr_noreturn_exclusions): Make it extern.
* attribs.h (get_attribute_namespace): Declare.
* tree-inline.c (function_attribute_inlinable_p): Use
get_attribute_name.

* c-attribs.c (handle_noreturn_attribute): No longer static.
* c-common.h (handle_noreturn_attribute, attr_noreturn_exclusions):
Declare.
* c-format.c (check_function_format): Use get_attribute_name.

* decl.c (duplicate_decls): Give an error when a function is
declared [[noreturn]] after its first declaration.
* parser.c (cp_parser_std_attribute): Don't treat C++11 noreturn
attribute as equivalent to GNU's.
* tree.c (std_attribute_table): Add noreturn.

* g++.dg/warn/noreturn-8.C: New test.
* g++.dg/warn/noreturn-9.C: New test.
* g++.dg/warn/noreturn-10.C: New test.
* g++.dg/warn/noreturn-11.C: New test.

From-SVN: r272486

PR tree-optimization/90626 - fold strcmp(a, b) == 0 to zero when one string length is exact and the other is unequal

gcc/ChangeLog:

PR tree-optimization/90626
* tree-ssa-strlen.c (strxcmp_unequal): New function.
(handle_builtin_string_cmp): Call it.

gcc/testsuite/ChangeLog:

PR tree-optimization/90626
* gcc.dg/strlenopt-65.c: New test.
* gcc.dg/strlenopt-66.c: New test.
* gcc.dg/strlenopt.h (strcmp, strncmp): Declare.

From-SVN: r272485

[Darwin, specs] Tidy some more linker options.

pie, no-pie and rdynamic are driver options, we can process them in the
relevant place and drop them once dealt with.  There's no need to generate
a new header to process the "no_compact_unwind" which is applied on the
basis of the target system.

Support for the -pie, -no_pie and -no_compact_unwind options should ideally
be checked at configure time, however the status quo is to assert that linkers
capable of targeting the relevant systems support these options (i.e. we trust
that the user doesn't attempt to configure inappropriately).

TODO: check the availability of the linker opts in configure rather than
trusting to the user.

This will fix the fail of pie-7.c, which is a result of failing to handle the
no-pie driver option.

2019-06-19  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.h (DRIVER_SELF_SPECS): Add RDYNAMIC, DARWIN_PIE_SPEC
and DARWIN_NOPIE_SPEC.
(RDYNAMIC): New, modified from DARWIN_EXPORT_DYNAMIC.
(DARWIN_PIE_SPEC): Collate from darwin.h and darwin9.h.
(DARWIN_NOPIE_SPEC): Collate from darwin10.h.
(DARWIN_NOCOMPACT_UNWIND): New from darwin10.h
(DARWIN_EXPORT_DYNAMIC): Delete.
* config/darwin10.h (LINK_GCC_C_SEQUENCE_SPEC): Move no_compact_unwind
and pie options processing to  darwin.h.
* config/darwin9.h (DARWIN_PIE_SPEC): Move pie processing to darwin.h

From-SVN: r272484

PR translation/90156 - add linter check suggesting to replace %<%s%> with %qs

gcc/c-family/ChangeLog:

PR translation/90156
* c-format.c (function_format_info::format_type): Adjust type.
(function_format_info::is_raw): New member.
(decode_format_type): Adjust signature.  Handle "raw" diag attributes.
(decode_format_attr): Adjust call to decode_format_type.
Avoid a redundant call to convert_format_name_to_system_name.
Avoid abbreviating the word "arguments" in a diagnostic.
(format_warning_substr): New function.
(avoid_dollar_number): Quote dollar sign in a diagnostic.
(finish_dollar_format_checking): Same.
(check_format_info): Same.
(struct baltoks_t): New.
(c_opers, c_keywords, cxx_keywords, badwords, contrs): New arrays.
(maybe_diag_unbalanced_tokens, check_tokens, check_plain): New
functions.
(check_format_info_main): Call check_plain.  Use baltoks_t.  Call
maybe_diag_unbalanced_tokens.
(handle_format_attribute): Spell out the word "arguments" in
a diagnostic.

gcc/testsuite/ChangeLog:

PR translation/90156
* gcc.dg/format/gcc_diag-11.c: Enable.

From-SVN: r272483

re PR fortran/69499 ([F03] ICE-on-invalid on combining select type with wrong statement)

2019-06-19  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/69499
* match.c (gfc_match_select_type):  SELECT TYPE is an executable
statement, and cannot appear in MODULE or SUBMODULE scope.

2019-06-19  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/69499
* gfortran.dg/pr69499.f90: New test.
* gfortran.dg/module_error_1.f90: Update dg-error string.

From-SVN: r272482

re PR fortran/69398 ([OOP] ICE on class with duplicate dimension attribute specified)

2019-06-19  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/69398
* decl.c (attr_decl): Check for duplicate DIMENSION attribute for a
CLASS entity.

2019-06-19  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/69398
* gfortran.dg/pr69398.f90: New test.

From-SVN: r272481

re PR fortran/87907 (ICE in resolve_contained_fntype, at fortran/resolve.c:587)

2019-06-19  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/87907
* resolve.c (resolve_contained_fntype): Do not dereference a NULL
pointer.

2019-06-19  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/87907
* gfortran.dg/pr87907.f90: New testcase.

From-SVN: r272480

[Darwin] Fix two off-by-one errors in the driver.

2019-06-19  Iain Sandoe  <iain@sandoe.co.uk>

        * config/darwin-driver.c (darwin_driver_init): Fix off-by-one errors
        in computing the number of options to be moved.

From-SVN: r272479

netbsd-elf.h (SYSARCH_ARM_SYNC_ICACHE): New definition.

*  config/arm/netbsd-elf.h (SYSARCH_ARM_SYNC_ICACHE): New definition.
(CLEAR_INSN_CACHE) Use it.

From-SVN: r272478

Fix value category bugs in std::reduce

* include/std/numeric (reduce(Iter, Iter, T, BinOp)): Fix value
category used in invocable check.
(reduce(Iter, Iter, T)): Pass initial value as rvalue.
* testsuite/26_numerics/reduce/2.cc: New test.

From-SVN: r272477

compiler: optimize string concatenations

    runtime.concatstring{2,3,4,5} are just wrappers of concatstrings.
    These wrappers don't provide any benefit, at least in the C
    calling convention we use, where passing arrays by value isn't an
    efficient thing. Change it to always use concatstrings.

    Also, the cap field of the slice passed to concatstrings is not
    necessary. So change it to pass a pointer and a length directly,
    which is more efficient than passing a slice header by value.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/182539

From-SVN: r272476

i386.md (cmpstrnsi): Remove dead code.

* config/i386/i386.md (cmpstrnsi): Remove dead code.

From-SVN: r272474

Simplify setjmp and non-local goto implementation (PR84521)

This fixes and simplifies the setjmp and non-local goto implementation.
Currently the virtual frame pointer is saved when using __builtin_setjmp or
a non-local goto.  Depending on whether a frame pointer is used, this may
either save SP or FP with an immediate offset.  However the goto or longjmp
always updates the hard frame pointer.

A receiver veneer in the original function then assigns the hard frame pointer
to the virtual frame pointer, which should, if it works correctly, again assign
SP or FP.  However the special elimination code in eliminate_regs_in_insn
doesn't do this correctly unless the frame pointer is used, and even if it
worked by writing SP, the frame pointer would still be corrupted.

A much simpler implementation is to always save and restore the hard frame
pointer.  This avoids 2 redundant instructions which add/subtract the virtual
frame offset.  A large amount of code can be removed as a result, including all
implementations of TARGET_BUILTIN_SETJMP_FRAME_VALUE (all of which already use
the hard frame pointer).  The expansion of nonlocal_goto on PA can be simplied
to just restore the hard frame pointer.

This fixes the most obvious issues, however there are still issues on targets
which define HARD_FRAME_POINTER_IS_FRAME_POINTER (arm, mips).
Each function could have a different hard frame pointer, so a non-local goto
may restore the wrong frame pointer (TARGET_BUILTIN_SETJMP_FRAME_VALUE could
be useful for this).

The i386 TARGET_BUILTIN_SETJMP_FRAME_VALUE was incorrect: if stack_realign_fp
is true, it would save the hard frame pointer value but restore the virtual
frame pointer which according to ix86_initial_elimination_offset can have a
non-zero offset from the hard frame pointer.

The ia64 implementation of nonlocal_goto seems incorrect since the helper
function moves the the frame pointer value into the static chain register
(so this patch does nothing to make it better or worse).

AArch64 + x86-64 bootstrap OK, new test passes on AArch64, x86-64 and Arm.

gcc/
PR middle-end/84521
* builtins.c (expand_builtin_setjmp_setup): Save
hard_frame_pointer_rtx.
(expand_builtin_setjmp_receiver): Do not emit sfp = fp move since we
restore fp.
* function.c (expand_function_start): Save hard_frame_pointer_rtx for
non-local goto.
* lra-eliminations.c (eliminate_regs_in_insn): Remove sfp = fp
elimination code.
(remove_reg_equal_offset_note): Remove unused function.
* reload1.c (eliminate_regs_in_insn): Remove sfp = hfp elimination
code.
* config/arc/arc.c (TARGET_BUILTIN_SETJMP_FRAME_VALUE): Remove.
(arc_builtin_setjmp_frame_value): Remove function.
* config/avr/avr.c  (TARGET_BUILTIN_SETJMP_FRAME_VALUE): Remove.
(avr_builtin_setjmp_frame_value): Remove function.
* config/i386/i386.c (TARGET_BUILTIN_SETJMP_FRAME_VALUE): Remove.
(ix86_builtin_setjmp_frame_value): Remove function.
* config/pa/pa.md (nonlocal_goto): Remove FP adjustment.
* config/sparc/sparc.c (TARGET_BUILTIN_SETJMP_FRAME_VALUE): Remove.
(sparc_builtin_setjmp_frame_value): Remove function.
* config/vax/vax.c (TARGET_BUILTIN_SETJMP_FRAME_VALUE): Remove.
(vax_builtin_setjmp_frame_value): Remove function.
* config/xtensa/xtensa.c (xtensa_frame_pointer_required): Force frame
pointer if has_nonlocal_label.

testsuite/
PR middle-end/84521
* gcc.c-torture/execute/pr84521.c: New test.

From-SVN: r272473

md.texi: Document vec_shl_<mode> pattern.

* doc/md.texi: Document vec_shl_<mode> pattern.
* optabs.def (vec_shl_optab): New optab.
* optabs.c (shift_amt_for_vec_perm_mask): Add shift_optab
argument, if == vec_shl_optab, check for left whole vector shift
pattern rather than right shift.
(expand_vec_perm_const): Add vec_shl_optab support.
* optabs-query.c (can_vec_perm_var_p): Mention also vec_shl optab
in the comment.
* tree-vect-generic.c (lower_vec_perm): Support permutations which
can be handled by vec_shl_optab.
* tree-vect-stmts.c (scan_store_can_perm_p): New function.
(check_scan_store): Use it.
(vectorizable_scan_store): If target can't do normal permutations,
try to use whole vector left shifts and if needed a VEC_COND_EXPR
after it.
* config/i386/sse.md (vec_shl_<mode>): New expander.

* gcc.dg/vect/vect-simd-8.c: If main is defined, don't include
tree-vect.h nor call check_vect.
* gcc.dg/vect/vect-simd-9.c: Likewise.
* gcc.dg/vect/vect-simd-10.c: New test.
* gcc.target/i386/sse2-vect-simd-8.c: New test.
* gcc.target/i386/sse2-vect-simd-9.c: New test.
* gcc.target/i386/sse2-vect-simd-10.c: New test.
* gcc.target/i386/avx2-vect-simd-8.c: New test.
* gcc.target/i386/avx2-vect-simd-9.c: New test.
* gcc.target/i386/avx2-vect-simd-10.c: New test.
* gcc.target/i386/avx512f-vect-simd-8.c: New test.
* gcc.target/i386/avx512f-vect-simd-9.c: New test.
* gcc.target/i386/avx512f-vect-simd-10.c: New test.

From-SVN: r272472

omp-low.c (lower_rec_input_clauses): Handle references properly in inscan clauses.

* omp-low.c (lower_rec_input_clauses): Handle references properly
in inscan clauses.
(lower_omp_scan): Likewise.
cp/
* cp-gimplify.c (cp_genericize_r): Handle OMP_CLAUSE_{IN,EX}CLUSIVE
like OMP_CLAUSE_SHARED.
testsuite/
* g++.dg/vect/simd-3.cc: New test.
* g++.dg/vect/simd-4.cc: New test.
* g++.dg/vect/simd-5.cc: New test.

From-SVN: r272471

* g++.dg/ubsan/pr63956.C: Adjust expected diagnostics.

From-SVN: r272470

Add new micro-benchmark for string operations.

2019-06-19  Martin Liska  <mliska@suse.cz>

* bench-stringop: New file.

From-SVN: r272469

compiler: stack allocate a buffer for non-escaping string ops

    For string concatenation, string to/from byte or rune slice
    conversion, and int to string conversion, if the result does not
    escape, we can allocate a small (32-element, or 4-byte for int to
    string) buffer on stack, and pass it to the runtime function. If
    the result fits in the buffer, it doesn't need to do a heap
    allocation.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/182538

From-SVN: r272468

re PR fortran/89103 (Allow blank format items in format strings)

2019-06-19  Jim MacArthur  <jim.macarthur@codethink.co.uk>
    Mark Eggleston  <mark.eggleston@codethink.com>

PR fortran/89103
* gfortran.texi: Add -fdec-blank-format-item
* invoke.texi: Add option to list of options.
* invoke.texi: Add to section on Commas in FORMAT specifications.
* io.c (check_format): At FMT_RPAREN goto finished if
-fdec-blank-format-item otherwise set error string.
* lang.opt: Add new option.
* options.c (set_dec_flags): Add SET_BITFLAG for
flag_dec_format_defaults.

* gfortran.dg/dec_format_empty_item_1.f: New test.
* gfortran.dg/dec_format_empty_item_2.f: New test.
* gfortran.dg/dec_format_empty_item_3.f: New test.

Co-Authored-By: Mark Eggleston <mark.eggleston@codethink.com>
From-SVN: r272467

pr88834.c: Move from here...

gcc/testsuite/ChangeLog:

2019-06-19  Kugan Vivekanandarajah  <kugan.vivekanandarajah@linaro.org>

* gcc.target/aarch64/pr88834.c: Move from here...
* gcc.target/aarch64/sve/pr88834.c: ...to here.

From-SVN: r272466

tree-ssa-address.c (preferred_mem_scale_factor): Handle when mem_mode is BLKmode.

gcc/ChangeLog:

2019-06-19  Kugan Vivekanandarajah  <kugan.vivekanandarajah@linaro.org>

* tree-ssa-address.c (preferred_mem_scale_factor): Handle when
mem_mode is BLKmode.

From-SVN: r272465

Daily bump.

From-SVN: r272464

compiler: avoid copy for string([]byte) conversion used in string concatenation

    If a string([]byte) conversion is used immediately in a string
    concatenation, we don't need to copy the backing store of the
    byte slice, as the runtime function doesn't hold any reference
    to it.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/182437

* go.dg/concatstring.go: New test.

From-SVN: r272460

Implement new serial algorithms from Parallelism TS (P0024R2)

These new (non-parallel) algorithms were added to C++17 along with the
parallel algorithms, but were missing from libstdc++.

* include/bits/algorithmfwd.h: Change title of doc group.
* include/bits/stl_algo.h (for_each_n): Add new C++17 algorithm from
P0024R2.
* include/bits/stl_numeric.h: Define doc group and add algos to it.
* include/std/numeric (__is_random_access_iter): New internal trait.
(reduce, transform_reduce, exclusive_scan, inclusive_scan)
(transform_exclusive_scan, transform_inclusive_scan): Likewise.
* testsuite/25_algorithms/for_each/for_each_n.cc: New test.
* testsuite/26_numerics/exclusive_scan/1.cc: New test.
* testsuite/26_numerics/inclusive_scan/1.cc: New test.
* testsuite/26_numerics/reduce/1.cc: New test.
* testsuite/26_numerics/transform_exclusive_scan/1.cc: New test.
* testsuite/26_numerics/transform_inclusive_scan/1.cc: New test.
* testsuite/26_numerics/transform_reduce/1.cc: New test.
* testsuite/util/testsuite_iterators.h (test_container::size()): New
member function.

From-SVN: r272459

xtensa: fix PR target/90922

Stack pointer adjustment code in prologue missed a case of no
callee-saved registers and a stack frame size bigger than 128 bytes.
Handle that case.

This fixes the following gcc tests with call0 ABI:
  gcc.c-torture/execute/stdarg-2.c
  gcc.dg/torture/pr55882.c
  gcc.dg/torture/pr57569.c

2019-06-18  Max Filippov  <jcmvbkbc@gmail.com>
gcc/
* config/xtensa/xtensa.c (xtensa_expand_prologue): Add stack
pointer adjustment for the case of no callee-saved registers and
stack frame bigger than 128 bytes.

From-SVN: r272455

[PR90921] Fortran OpenACC 'declare' directive's module handling causes duplicate data clauses

gcc/fortran/
PR fortran/90921
* trans-decl.c (finish_oacc_declare): Reset module_oacc_clauses
before scanning each namespace.
gcc/testsuite/
PR fortran/90921
* gfortran.dg/goacc/declare-3.f95: Update.

From-SVN: r272454

[PR85221] Set 'omp declare target', 'omp declare target link' attributes for Fortran OpenACC 'declare'd variables

gcc/fortran/
PR fortran/85221
* trans-decl.c (add_attributes_to_decl): Handle OpenACC 'declare'
directive.
gcc/testsuite/
PR fortran/85221
* gfortran.dg/goacc/declare-3.f95: New file.

From-SVN: r272453

[PR90859] Document status quo for "[OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'"

gcc/testsuite/
PR middle-end/90859
* c-c++-common/goacc/firstprivate-mappings-1.c: Update.

From-SVN: r272452

Test cases to verify OpenACC 'firstprivate' mappings

gcc/testsuite/
* c-c++-common/goacc/firstprivate-mappings-1.c: New file.
* g++.dg/goacc/firstprivate-mappings-1.C: Likewise.
libgomp/
* testsuite/libgomp.oacc-c++/firstprivate-mappings-1.C: New file.
* testsuite/libgomp.oacc-c-c++-common/firstprivate-mappings-1.c:
Likewise.

From-SVN: r272451

Fix description of 'GOMP_MAP_FIRSTPRIVATE'

..., which got garbled in r230275.

include/
* gomp-constants.h (enum gomp_map_kind): Fix description of
'GOMP_MAP_FIRSTPRIVATE'.

From-SVN: r272450

Add missing results check in 'libgomp.fortran/allocatable3.f90'

libgomp/
* testsuite/libgomp.fortran/allocatable3.f90: Add missing results
check.

From-SVN: r272449

Add 'libgomp.oacc-fortran/allocatable-array-1.f90'

libgomp/
* testsuite/libgomp.oacc-fortran/allocatable-array-1.f90: New
file.

From-SVN: r272448

[PR90743] Fortran 'allocatable' with OpenACC data/OpenMP 'target' 'map' clauses

Test what OpenMP 5.0 has to say on this topic.  And, do the same for OpenACC.

libgomp/
PR fortran/90743
* oacc-parallel.c (GOACC_parallel_keyed): Handle NULL mapping
case.
* testsuite/libgomp.fortran/target-allocatable-1-1.f90: New file.
* testsuite/libgomp.fortran/target-allocatable-1-2.f90: Likewise.
* testsuite/libgomp.oacc-fortran/allocatable-1-1.f90: Likewise.
* testsuite/libgomp.oacc-fortran/allocatable-1-2.f90: Likewise.

From-SVN: r272447

[PR90861] Document status quo for OpenACC 'declare' not cleaning up for VLAs

gcc/testsuite/
PR testsuite/90861
* c-c++-common/goacc/declare-pr90861.c: New file.
libgomp/
PR testsuite/90861
* testsuite/libgomp.oacc-c-c++-common/declare-vla.c: Update.

From-SVN: r272446

[PR90868] Document status quo for duplicate OpenACC 'declare' directives for 'extern' variables

gcc/testsuite/
PR testsuite/90868
* c-c++-common/goacc/declare-1.c: Update.
* c-c++-common/goacc/declare-2.c: Likewise.

From-SVN: r272445

[PR90862] OpenACC 'declare' ICE when nested inside another construct

gcc/
PR middle-end/90862
* omp-low.c (check_omp_nesting_restrictions): Handle
GF_OMP_TARGET_KIND_OACC_DECLARE.
gcc/testsuite/
PR middle-end/90862
* c-c++-common/goacc/declare-1.c: Update.
* c-c++-common/goacc/declare-2.c: Likewise.
libgomp/
PR middle-end/90862
* testsuite/libgomp.oacc-c-c++-common/declare-1.c: Update.

From-SVN: r272444

re PR c++/84698 (ICE when using noexcept(noexcept()) declaration on global friend function of template class)

PR c++/84698
* g++.dg/cpp0x/noexcept42.C: New test.

From-SVN: r272443

re PR c++/71548 (Invalid declaration involving template template param causes crash)

PR c++/71548
* g++.dg/cpp0x/variadic177.C: New test.

From-SVN: r272438

i386.md (@cmp<mode>_1): Rename from cmp<mode>_1.

* config/i386/i386.md (@cmp<mode>_1): Rename from cmp<mode>_1.
(@add<mode>3_carry): Rename from add<mode>3_carry.
(@sub<mode>3_carry_ccc): Rename from sub<mode>3_carry_ccc.
(@sub<mode>3_carry_ccgz): Rename form sub<mode>3_carry_ccgz.
(@copysign<mode>3_const): Rename from copysign<mode>3_const.
(@copysign<mode>3_var): Rename from copysign<mode>3_var.
(@xorsign<mode>3_1): Rename from xorsign<mode>3_1.
(@x86_shift<mode>_adj_1): Rename from x86_shift<mode>_adj_1.
(@x86_shift<mode>_adj_2): Rename from x86_shift<mode>_adj_2.
(@x86_shift<mode>_adj_3): Rename from x86_shift<mode>_adj_3.
(cmpstrnsi): Use gen_cmp_1.
(lwp_slwpcb): Use gen_lwp_slwpcb_1.
(@lwp_slwpcb<mode>_1): Rename from lwp_slwpcb<mode>_1.
(@umonitor_<mode>): Rename from umonitor_<mode>.
* config/i386/i386-expand.c (ix86_expand_copysign):
Use gen_copysign3_const and gen_copysign3_var.
(ix86_expand_xorsign): Use gen_xorsign3_1.
(ix86_expand_branch): Use gen_sub3_carry_ccc,
gen_sub3_carry_ccgz and gen_cmp1.
(ix86_expand_int_addcc): Use gen_sub3_carry and gen_add3_carry.
(ix86_split_ashl): Use gen_x86_shift_adj_1 and gen_x86_shift_adj_2.
(ix86_split_ashr): Use gen_x86_shift_adj_1 and gen_x86_shift_adj_3.
(ix86_split_lshr): Ditto.
(ix86_expand_builtin) <case IX86_BUILTIN_UMONITOR>: Use gen_umonitor.

From-SVN: r272432

* constexpr.c (cxx_eval_store_expression): Delay target evaluation.

From-SVN: r272431

Consolidate constexpr array handling.

* constexpr.c (eval_and_check_array_index): Split out from...
(cxx_eval_array_reference): ...here.
(cxx_eval_store_expression): Use it here, too.
(diag_array_subscript): Take location.  Strip location wrapper.

From-SVN: r272430

Handle constexpr conversion from and then to the same type.

* constexpr.c (cxx_eval_constant_expression): Handle conversion from
and then to the same type.

From-SVN: r272429

tree.c (build_constructor): Add MEM_STAT_DECL.

* tree.c (build_constructor): Add MEM_STAT_DECL.

gcc/cp/
* constexpr.c (unshare_constructor): Add MEM_STAT_DECL.

From-SVN: r272428

[AArch64] Add a new CC mode for SVE conditions

The SVE ACLE patches need to introduce a new CC_NZC mode for the
conditions that can be tested after a PTRUE.  In particular, LT needs
to map to "mi"/"first" and GE to "pl"/"nfrst", instead of the normal
CC mapping.

Another advantage of using a separate mode is that we can print the SVE
names of the conditions, which makes the output a bit easier to read.
It therefore seems like an independent improvement that can go in now.

The patch also avoids using (compare X (const_int 0)), because that gets
folded away when used with LTU and GEU ("cc"/"last" and "cs"/"nlast").
Just using an unspec should be OK.

The full set of conditions can't be tested without other SVE ACLE patches.

2019-06-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-modes.def (CC_NZC): New CC_MODE.
* config/aarch64/aarch64-sve.md (*<optab><mode>3_cc)
(ptest_ptrue<mode>, while_ult<GPI:mode><PRED_ALL:mode>)
(*while_ult<GPI:mode><PRED_ALL:mode>_cc, *cmp<cmp_op><mode>)
(*cmp<cmp_op><mode>_ptest, *cmp<cmp_op><mode>_cc)
(*pred_cmp<cmp_op><mode>_combine, *pred_cmp<cmp_op><mode>)
(vec_cmp<mode><vpred>, vec_cmpu<mode><vpred>, cbranch<mode>4):
Use CC_NZC instead of CC.
* config/aarch64/aarch64.md (condjump): Print a '.' in SVE conditions.
* config/aarch64/aarch64.c (aarch64_sve_condition_codes): New variable.
(aarch64_print_operand): Handle E_CC_NZCmode.
(aarch64_emit_sve_ptrue_op_cc): Use gen_set_clobber_cc_nzc instead
of gen_set_clobber_cc.

gcc/testsuite/
* gcc.target/aarch64/sve/struct_vect_18.c: Allow branches to
contain dots.
* gcc.target/aarch64/sve/struct_vect_19.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_20.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_21.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_22.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_23.c: Likewise.
* gcc.target/aarch64/sve/unroll-1.c: Likewise.
* gcc.target/aarch64/sve/while_1.c: Check for b.any.

From-SVN: r272427

[AArch64] Tabify aarch64-sve.md

2019-06-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-sve.md: Tabify file.

From-SVN: r272426

[AArch64] Factor out pfalse predicate creation

Following on from the previous ptrue patch.

2019-06-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-protos.h (aarch64_pfalse_reg): Declare.
* config/aarch64/aarch64.c (aarch64_pfalse_reg): New function.
* config/aarch64/aarch64-sve.md: Use it.

From-SVN: r272425

[AArch64] Factor out ptrue predicate creation

This is the first step to canonicalising predicate constants so that
they can be reused between modes.

2019-06-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-protos.h (aarch64_ptrue_reg): Declare.
* config/aarch64/aarch64.c (aarch64_ptrue_reg): New functions.
(aarch64_expand_sve_widened_duplicate, aarch64_expand_sve_mem_move)
(aarch64_maybe_expand_sve_subreg_move, aarch64_evpc_rev_local)
(aarch64_expand_sve_vec_cmp_int): Use it.
(aarch64_expand_sve_vec_cmp_float): Likewise.
* config/aarch64/aarch64-sve.md: Likewise throughout.

From-SVN: r272424

[AArch64] Simplify SVE IFN_COND patterns

This patch makes the binary IFN_COND patterns use the same approach
as the ternary patterns, with one pattern handling the cases in
which the "else" value isn't tied to one of the other inputs.

2019-06-18  Richard Sandiford  <richard.sandiford@arm.com>
    Kugan Vivekanandarajah  <kuganv@linaro.org>

gcc/
* config/aarch64/aarch64-sve.md (*cond_<optab><mode>_0): Delete.
(*cond_<optab><mode>_z): Fold into...
(*cond_<optab><mode>_any): ...here.  Also handle cases in which
operand 4 can be tied to operand 0 (either inherently or via RA).

From-SVN: r272423

pr81563.c (dg-final): Check that no registers are restored from %esp.

* gcc.target/i386/pr81563.c (dg-final): Check that no
registers are restored from %esp.

From-SVN: r272422

re PR debug/90900 (ICE in copy_rtx, at rtl.c:376)

2019-06-18  Richard Biener  <rguenther@suse.de>

PR debug/90900
* cfgexpand.c (expand_debug_expr): Treat NOTE_P DECL_RTL
as if optimized away.

* gcc.dg/gomp/pr90900.c: New testcase.

From-SVN: r272421