Catch division by zero errors in array sizes.

2019-12-30  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/92961
* gfortran.h (gfc_seen_div0): Add declaration.
* arith.h (gfc_seen_div0): Add definition.
(eval_intrinsic): For integer division by zero, set gfc_seen_div0.
* decl.c (variable_decl):  If resolution resp. simplification
fails for array spec and a division of zero error has been
seen, return MATCH_ERROR.

2019-12-30  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/92961
* gfortran.dg/arith_divide_2.f90: New test.

From-SVN: r279762

Daily bump.

From-SVN: r279761

re PR c++/92745 (Initializing array with vec4 results in compile error)

PR c++/92745
* g++.dg/cpp0x/initlist118.C: Add -Wno-psabi -w to dg-options.

From-SVN: r279758

PR c++/88337 - Implement P1327R1: Allow dynamic_cast in constexpr.

This patch implements
<http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1327r1.html>.

When build_dynamic_cast realizes that a dynamic_cast needs a run-time check, it
generates a call to __dynamic_cast -- see dyncast.cc in libsupc++ for its
definition.  The gist of my approach is to evaluate such a call at compile time.

* constexpr.c (cxx_dynamic_cast_fn_p): New function.
(extract_obj_from_addr_offset): New function.
(get_component_with_type): New function.
(cxx_eval_dynamic_cast_fn): New function.
(cxx_eval_call_expression): Call cxx_eval_dynamic_cast_fn for a call
to __dynamic_cast.
(potential_constant_expression_1): Don't give up on
cxx_dynamic_cast_fn_p.
* rtti.c (build_dynamic_cast_1): When creating a call to
__dynamic_cast, use the location of the original expression.

* g++.dg/cpp2a/constexpr-dynamic1.C: New test.
* g++.dg/cpp2a/constexpr-dynamic10.C: New test.
* g++.dg/cpp2a/constexpr-dynamic11.C: New test.
* g++.dg/cpp2a/constexpr-dynamic12.C: New test.
* g++.dg/cpp2a/constexpr-dynamic13.C: New test.
* g++.dg/cpp2a/constexpr-dynamic14.C: New test.
* g++.dg/cpp2a/constexpr-dynamic15.C: New test.
* g++.dg/cpp2a/constexpr-dynamic16.C: New test.
* g++.dg/cpp2a/constexpr-dynamic17.C: New test.
* g++.dg/cpp2a/constexpr-dynamic2.C: New test.
* g++.dg/cpp2a/constexpr-dynamic3.C: New test.
* g++.dg/cpp2a/constexpr-dynamic4.C: New test.
* g++.dg/cpp2a/constexpr-dynamic5.C: New test.
* g++.dg/cpp2a/constexpr-dynamic6.C: New test.
* g++.dg/cpp2a/constexpr-dynamic7.C: New test.
* g++.dg/cpp2a/constexpr-dynamic8.C: New test.
* g++.dg/cpp2a/constexpr-dynamic9.C: New test.

From-SVN: r279755

re PR target/93078 (Missing fma and round functions auto-vectorization with x86-64 (sse2))

PR target/93078
* config/i386/i386-builtins.c (ix86_builtin_vectorized_function):
Remove CASE_CFN_RINT handling.
* config/i386/i386-builtin.def (IX86_BUILTIN_RINTPD,
IX86_BUILTIN_RINTPS, IX86_BUILTIN_RINTPD256, IX86_BUILTIN_RINTPS256):
Remove.
* config/i386/sse.md (nearbyint<mode>2, rint<mode>2): New expanders
with VF iterator.

* gcc.target/i386/sse4_1-pr93078.c: New test.
* gcc.target/i386/avx-pr93078.c: New test.
* gcc.target/i386/avx512f-pr93078.c: New test.

From-SVN: r279754

Unshare DR_STEP before gimplifying it

In this testcase we use an unmasked SVE loop with an Advanced SIMD
epilogue (because we don't yet support fully-masked downward loops).
The main loop uses a gather load for the strided access while the
epilogue loop builds the access from scalars instead.  In both cases
we gimplify expressions based on the DR_STEP and insert them in the
loop preheader.

The problem was that the gather load code didn't copy the DR_STEP before
gimplifying it, meaning that the epilogue loop tried to reuse a result
from the (non-dominating) main loop preheader.

It looks at first glance like there could be other instances of this too,
but this patch just deals with the gather/scatter case.

2019-12-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (vect_get_strided_load_store_ops): Copy
DR_STEP before gimplifying it.

gcc/testsuite/
* gcc.dg/vect/vect-strided-epilogue-1.c: New test.

From-SVN: r279753

Check for a supported comparison when using EXTRACT_LAST_REDUCTION

The EXTRACT_LAST_REDUCTION handling needs to generate a separate
comparison instruction that feeds the vector mask argument of the
IFN_EXTRACT_LAST call.  We weren't checking whether that comparison
was supported, leading to an ICE on the testcase.

2019-12-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (vectorizable_condition): For extract-last
reductions, check that the target supports the required comparison
operation.

gcc/testsuite/
* gcc.dg/vect/vect-cond-12.c: New test.

From-SVN: r279752

Daily bump.

From-SVN: r279751

re PR bootstrap/93074 (build FAIL with --enable-offload-targets=nvptx-none)

PR bootstrap/93074
* plugin/cuda/cuda.h (cuDeviceGetName, cuDriverGetVersion): Declare.
(cuDeviceTotalMem, cuMemGetInfo): Likewise.  Define to *_v2.

From-SVN: r279747

Daily bump.

From-SVN: r279746

[AArch64] Fix typo in V_INT_CONTAINER

All VNx2 V_INT_CONTAINER entries should map to VNx2DI.  The lower-case
version was already correct.

2019-12-27  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/iterators.md (V_INT_CONTAINER): Fix VNx2SF entry.

gcc/testsuite/
* gcc.target/aarch64/sve/mixed_size_11.c: New test.

From-SVN: r279743

Add missing target check for fully-masked fold-left reductions

The fold-left reduction code has a (rarely-used) fallback that handles
cases in which the loop is fully-masked and the target has no native
support for the reduction.  The fallback includea a VEC_COND_EXPR
between the reduction vector and a safe value, so we should check
whether that VEC_COND_EXPR is supported.

2019-12-27  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-loop.c (vectorizable_reduction): Check whether the
target supports the required VEC_COND_EXPR operation before
allowing the fallback handling of masked fold-left reductions.

gcc/testsuite/
* gcc.target/aarch64/sve/mixed_size_10.c: New test.

From-SVN: r279742

Daily bump.

From-SVN: r279739

re PR c++/92438 (Function declaration parsed incorrectly with `-std=c++1z`)

PR c++/92438
* parser.c (cp_parser_constructor_declarator_p): If open paren
is followed by RID_ATTRIBUTE, skip over the attribute tokens and
try to parse type specifier.

* g++.dg/ext/attrib61.C: New test.

From-SVN: r279736

Daily bump.

From-SVN: r279735

Define HAVE_ for math long double functions declared in vxworks headers

When cross-building for vxworks, test for declarations of long double
functions in math.h.  We don't normally test for these functions when
cross compiling, because link tests don't work, or ever really, but
not defining them as available causes replacements to be defined in
ways that may cause duplicate definition linker errors if the units
defining both the replacement and the actual implementation are
brought in because of other symbols.

for libstdc++-v3/ChangeLog

* crossconfig.m4 (GLIBCXX_CROSSCONFIG) [*-vxworks*]: Define
long double functions as available if declared by math.h.
(GLIBCXX_CHECK_MATH_DECL, GLIBCXX_CHECK_MATH_DECLS): New.
* configure: Rebuild.

From-SVN: r279731

Daily bump.

From-SVN: r279730

rs6000: re-enable web and rnreg with -funroll-loops

Previously, limited unrolling was enabled at O2 for powerpc in r278034.  At that
time, -fweb and -frename-registers were not enabled together with -funroll-loops
even for -O3.  After that, we notice there are some performance degradations on
SPEC2006fp which caused by without web and rnreg.  This patch enable -fweb
and -frename-registers for -funroll-loops to align original behavior before
r278034.

gcc/
2019-12-23  Jiufu Guo  <guojiufu@linux.ibm.com>

* config/rs6000/rs6000.c (rs6000_option_override_internal): Enable
-fweb and -frename-registers with -funroll-loops

From-SVN: r279725

re PR go/93020 (Final patches to build gcc-10 on GNU/Hurd)

PR go/93020
    libgo: Hurd portability patches

    By Svante Signell.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/212409

From-SVN: r279724

Daily bump.

From-SVN: r279723

Restrict 'c-c++-common/goacc/mdc-1.c' to LP64, LLP64

The tree dump scanning has certain expectations.

gcc/testsuite/
* c-c++-common/goacc/mdc-1.c: Restrict to LP64, LLP64.

From-SVN: r279720

[C++] Make same_type_p return false for gnu_vector_type_p differences (PR 92789)

As Jason pointed out in the review of the C++ gnu_vector_type_p patch:

    https://gcc.gnu.org/ml/gcc-patches/2019-12/msg00173.html

the real fix for the XFAILs in acle/general-c++/gnu_vectors_*.C is to
make same_type_p return false for two types if one is gnu_vector_type_p
and the other isn't.  This patch does that and fixes the fallout.

Originally I'd tried to make it so that "X *" and "Y *" are
interconvertible whenever X and Y are, and similarly for
"X &" and "Y &".  That doesn't fall out naturally though,
and is different from how -flax-vector-conversions works.
The patch therefore accepts all the consequences of making
X and Y !same_type_p instead of trying to work around them.

2019-12-23  Richard Sandiford  <richard.sandiford@arm.com>

gcc/cp/
PR c++/92789
* typeck.c (structural_comptypes): Make sure that two vector types
agree on gnu_vector_type_p.

gcc/testsuite/
PR c++/92789
* g++.dg/ext/sve-sizeless-2.C (statements): Expect pointer
difference and comparisons between GNU and non-GNU types
to be rejected.  Expect __is_same to be false for such pairs.
* g++.target/aarch64/sve/acle/general-c++/gnu_vectors_1.C: Remove
XFAILs.  Expect conversions between SVE vector pointers and
GNU vector pointers to be rejected.  Test references.
* g++.target/aarch64/sve/acle/general-c++/gnu_vectors_2.C: Likewise.

From-SVN: r279717

[C++] Fix ICE for binding lax vector conversions to references (PR 93014)

This test:

typedef unsigned int v4si __attribute__ ((vector_size(16)));
typedef unsigned char v16qi __attribute__ ((vector_size(16)));
extern v16qi x;
v4si &y = x;

ICEs with:

a.c:4:11: internal compiler error: in convert_like_real, at cp/call.c:7670

This started with r260780, which had the effect of making lvalue_kind
look through VIEW_CONVERT_EXPR in all cases, not just for location
wrappers.  This also means that:

typedef unsigned int v4si __attribute__ ((vector_size(16)));
typedef unsigned char v16qi __attribute__ ((vector_size(16)));
extern v16qi x;
v4si &y = reinterpret_cast<v4si>(x);

is now valid despite the result of the cast being an rvalue.

The patch attempts to fix that by calling rvalue on the input to the
conversion, so that the tree looks the same as for:

  extern v16qi x;
  v4si &y = (v4si)x;

which is already handled correctly.

2019-12-23  Richard Sandiford  <richard.sandiford@arm.com>

gcc/cp/
* cvt.c (ocp_convert): Apply rvalue to the source of vector
conversions.
* typeck.c (build_reinterpret_cast_1): Likewise.

gcc/testsuite/
* g++.dg/ext/vector39.C: New test.

From-SVN: r279716

Daily bump.

From-SVN: r279715

Add OpenACC 2.6 `acc_get_property' support

Add generic support for the OpenACC 2.6 `acc_get_property' and
`acc_get_property_string' routines, as well as full handlers for the
host and the NVPTX offload targets and minimal handlers for the HSA,
Intel MIC, and AMD GCN offload targets.

Included are C/C++ and Fortran tests that, in particular, print
the property values for acc_property_vendor, acc_property_memory,
acc_property_free_memory, acc_property_name, and acc_property_driver.
The output looks as follows:

Vendor: GNU
Name: GOMP
Total memory: 0
Free memory: 0
Driver: 1.0

with the host driver (where the memory related properties are not
supported for the host device and yield 0, conforming to the standard)
and output like:

Vendor: Nvidia
Total memory: 12651462656
Free memory: 12202737664
Name: TITAN V
Driver: CUDA Driver 9.1

with the NVPTX driver.

2019-12-22  Maciej W. Rozycki  <macro@codesourcery.com>
    Frederik Harwath  <frederik@codesourcery.com>
    Thomas Schwinge  <tschwinge@codesourcery.com>

include/
* gomp-constants.h (gomp_device_property): New enum.

libgomp/
* libgomp.h (gomp_device_descr): Add `get_property_func' member.
* libgomp-plugin.h (gomp_device_property_value): New union.
(gomp_device_property_value): New prototype.
* openacc.h (acc_device_t): Add `acc_device_current' enumeration
constant.
(acc_device_property_t): New enum.
(acc_get_property, acc_get_property_string): New prototypes.
* oacc-init.c (acc_get_device_type): Also assert that result
is not `acc_device_current'.
(get_property_any, acc_get_property, acc_get_property_string):
New functions.
* openacc.f90 (openacc_kinds): Add `acc_device_current' and
`acc_property_memory', `acc_property_free_memory',
`acc_property_name', `acc_property_vendor' and
`acc_property_driver' constants.  Add `acc_device_property' data
type.
(openacc_internal): Add `acc_get_property' and
`acc_get_property_string' interfaces.  Add `acc_get_property_h',
`acc_get_property_string_h', `acc_get_property_l' and
`acc_get_property_string_l'.
* oacc-host.c (host_get_property): New function.
(host_dispatch): Wire it.
* target.c (gomp_load_plugin_for_device): Handle `get_property'.
* libgomp.map (OACC_2.6): Add `acc_get_property', `acc_get_property_h_',
`acc_get_property_string' and `acc_get_property_string_h_' symbols.
* libgomp.texi (OpenACC Runtime Library Routines): Add
`acc_get_property'.
(acc_get_property): New node.
* plugin/plugin-gcn.c (GOMP_OFFLOAD_get_property): New
function (stub).
* plugin/plugin-hsa.c (GOMP_OFFLOAD_get_property): New function.
* plugin/plugin-nvptx.c (CUDA_CALLS): Add `cuDeviceGetName',
`cuDeviceTotalMem', `cuDriverGetVersion' and `cuMemGetInfo'
calls.
(GOMP_OFFLOAD_get_property): New function.
(struct ptx_device): Add new field "name".
(cuda_driver_version_s): Add new static variable ...
(nvptx_init): ... and init from here.

* testsuite/libgomp.oacc-c-c++-common/acc_get_property.c: New test.
* testsuite/libgomp.oacc-c-c++-common/acc_get_property-2.c: New test.
* testsuite/libgomp.oacc-c-c++-common/acc_get_property-3.c: New test.
* testsuite/libgomp.oacc-c-c++-common/acc_get_property-aux.c: New file
with test helper functions.

* testsuite/libgomp.oacc-fortran/acc_get_property.f90: New test.

liboffloadmic/
* plugin/libgomp-plugin-intelmic.cpp (GOMP_OFFLOAD_get_property):
New function.

Reviewed-by: Thomas Schwinge <thomas@codesourcery.com>
Co-Authored-By: Frederik Harwath <frederik@codesourcery.com>
Co-Authored-By: Thomas Schwinge <tschwinge@codesourcery.com>
From-SVN: r279710

Restrict some aarch64 testcases to little-endian

2019-12-21  Andrew Pinski  <apinski@marvell.com>

        PR testsuite/92998
        * gcc.target/aarch64/sve/acle/general/dupq_1.c:
        Restrict to aarch64_little_endian only.
        * gcc.target/aarch64/torture/simd-abi-8.c:
        Likewise.

From-SVN: r279709

libgomp/test: Fix compilation for build sysroot

Fix a problem with the libgomp testsuite using a method to determine
the compiler to use resulting in the tool being different from one the
library has been built with, and causing a catastrophic failure from the
lack of a suitable `--sysroot=' option where the `--with-build-sysroot='
configuration option has been used to build the compiler resulting in
the inability to link executables.

Address this problem by defining the compiler to use, via the
GCC_UNDER_TEST TCL variable, set in the DejaGNU configuration file from
$CC by autoconf, which will have all the required options set for the
target compiler to build executables in the environment configured,
removing failures like:

.../bin/riscv64-linux-gnu-ld: cannot find crt1.o: No such file or directory
.../bin/riscv64-linux-gnu-ld: cannot find -lm
.../bin/riscv64-linux-gnu-ld: cannot find -lpthread
.../bin/riscv64-linux-gnu-ld: cannot find -lc
.../bin/riscv64-linux-gnu-ld: cannot find -latomic
collect2: error: ld returned 1 exit status
compiler exited with status 1
FAIL: libgomp.c/../libgomp.c-c++-common/atomic-18.c (test for excess errors)
Excess errors:
.../bin/riscv64-linux-gnu-ld: cannot find crt1.o: No such file or directory
.../bin/riscv64-linux-gnu-ld: cannot find -lm
.../bin/riscv64-linux-gnu-ld: cannot find -lpthread
.../bin/riscv64-linux-gnu-ld: cannot find -lc
.../bin/riscv64-linux-gnu-ld: cannot find -latomic

UNRESOLVED: libgomp.c/../libgomp.c-c++-common/atomic-18.c compilation failed to produce executable

and bringing overall test results for the `riscv64-linux-gnu' target
(here with the `x86_64-linux-gnu' host and RISC-V QEMU in the Linux user
emulation mode as the target board) from:

=== libgomp Summary ===

# of expected passes 44
# of unexpected failures 3274
# of unresolved testcases 3241
# of unsupported tests 548

to:

=== libgomp Summary ===

# of expected passes 6834
# of unexpected failures 4
# of expected failures 4
# of unsupported tests 518

libgomp/
* testsuite/libgomp-test-support.exp.in (GCC_UNDER_TEST): New
variable.

From-SVN: r279708

* doc/invoke.texi (-flto): Use "compile time" as a noun.

From-SVN: r279707

testsuite: Fix run-time tracking down of `libgcc_s'

Fix a catastrophic libgo testsuite failure in cross-compilation where
the shared `libgcc_s' library cannot be found by the loader at run time
in build-tree testing and consequently all test cases fail the execution
stage, giving output (here with the `x86_64-linux-gnu' host and the
`riscv64-linux-gnu' target, with RISC-V QEMU in the Linux user emulation
mode as the target board) like:

spawn qemu-riscv64 -E LD_LIBRARY_PATH=.:.../riscv64-linux-gnu/lib64/lp64d/libgo/.libs ./a.exe
./a.exe: error while loading shared libraries: libgcc_s.so.1: cannot open shared object file: No such file or directory
FAIL: archive/tar

To do so rework `gcc-set-multilib-library-path' so as not to rely on the
`rootme' TCL variable to have been preset in testsuite invocation, which
only works for the GCC test suites and not for library test suites, and
also use `remote_exec host' rather than `exec' to invoke the compiler in
determination of `libgcc_s' locations, so that the solution works in
remote testing as well while also avoiding the hardcoded limit of the
executable's path length imposed by `exec'.

This is based on an observation that the multilib root directory can be
determined by stripping out the multilib directory in effect as printed
with the `-print-multi-directory' option from the path produced by the
`-print-file-name=' option.  And then individual full multilib paths can
be assembled for the other multilibs by appending their respective
multilib directories to the multilib root directory.

Unlike with the old solution the full multilib paths are not checked for
the presence of the shared `libgcc_s' library there, but that is
supposed to be harmless.  Also the full multilib path for the multilib
used with the compiler used for testing will now come first, which
should reduce run-time processing in the usual case.

With this change in place test output instead looks like:

spawn qemu-riscv64 -E LD_LIBRARY_PATH=.:.../riscv64-linux-gnu/lib64/lp64d/libgo/.libs:..././gcc/lib64/lp64d:..././gcc/.:..././gcc/lib32/ilp32:..././gcc/lib32/ilp32d:..././gcc/lib64/lp64 ./a.exe
PASS
PASS: archive/tar

No summary comparison, because the libgo testsuite does not provide one
in this configuration for some reason, however this change improves
overall results from 0 PASSes and 159 FAILs to 133 PASSes and 26 FAILs.

gcc/testsuite/
* lib/gcc-defs.exp (gcc-set-multilib-library-path): Use
`-print-file-name=' to determine the multilib root directory.
Use `remote_exec host' rather than `exec' to invoke the
compiler.

From-SVN: r279706

Daily bump.

From-SVN: r279705

libada: Fix shared library installation with `--disable-libada'

Provide a default value of $(toolexeclibdir) for $(ADA_RTL_DSO_DIR), so
that in a `--disable-libada' configuration `make install' places shared
gnatlib libraries, built with `make -C gcc gnatlib-shared', in their
intended version-specific location, fixing a commit r276424 ("libada:
Respect `--enable-version-specific-runtime-libs'") regression.

gcc/ada/
* gcc-interface/Makefile.in (toolexeclibdir): New variable.

From-SVN: r279702

[OMP] Restore 'omp declare target link' handling

    PASS: libgomp.c/target-link-1.c (test for excess errors)
    [-PASS:-]{+FAIL:+} libgomp.c/target-link-1.c execution test

We need to revert one line of code change from r279625.

libgomp/
* target.c (gomp_map_vars_internal): Restore 'omp declare target
link' handling.

From-SVN: r279701

[PR93026, PR92929] Adjust 'gfortran.dg/goacc/finalize-1.f' for r279626 changes

gcc/testsuite/
PR fortran/93026
PR middle-end/92929
* gfortran.dg/goacc/finalize-1.f: Adjust.

From-SVN: r279700

re PR fortran/91661 (ICE in gfc_conv_intrinsic_dot_product, at fortran/trans-intrinsic.c:4804)

2019-12-21  Harald Anlauf  <anlauf@gmx.de>

PR fortran/91661
* gfortran.dg/pr91661.f90: New test.

From-SVN: r279699

re PR fortran/92990 (INVALID code with NULLIFY – partially misleading error message "If bounds remapping is specified at (1), the pointer target shall not be NULL")

2019-12-21  Harald Anlauf  <anlauf@gmx.de>

PR fortran/92990
* match.c (gfc_match_nullify): Check for valid pointer object.
Reject bounds remapping.

PR fortran/92990
* gfortran.dg/pr92990.f90: New test.

From-SVN: r279698

re PR fortran/92753 (ICE in gfc_trans_call, at fortran/trans-stmt.c:392)

2019-12-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/92753
* expr.c (find_inquiry_ref): Catch INQUIRY_LEN case, where the
temporary expression has been converted to a constant and make
the new expression accordingly. Correct the error in INQUIRY_RE
and INQUIRY_IM cases. The original rather than the resolved
expression was being used as the source in mpfr_set.

2019-12-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/92753
* gfortran.dg/inquiry_type_ref_5.f90 : New test.

From-SVN: r279696

Avoid segfault when doing IPA-VRP but not IPA-CP (PR 93015)

2019-12-21  Martin Jambor  <mjambor@suse.cz>

PR ipa/93015
* ipa-cp.c (ipcp_store_vr_results): Check that info exists

testsuite/
* gcc.dg/lto/pr93015_0.c: New test.

From-SVN: r279695

Daily bump.

From-SVN: r279690

re PR middle-end/91512 (Fortran compile time regression.)

PR middle-end/91512
PR fortran/92738
* lang.opt (-finline-arg-packing): Add trailing dot to help text.

From-SVN: r279687

PR c++/92745 - bogus error when initializing array of vectors.

In r268428 I changed reshape_init_r in such a way that when it sees
a nested { } in a CONSTRUCTOR with missing braces, it just returns
the initializer:
+     else if (COMPOUND_LITERAL_P (stripped_init)
...
+         ++d->cur;
+         gcc_assert (!BRACE_ENCLOSED_INITIALIZER_P (stripped_init));
+         return init;

But as this test shows, that's incorrect: if TYPE is an array, we need
to proceed to reshape_init_array_1 which will iterate over the array
initializers:
 6006   /* Loop until there are no more initializers.  */
 6007   for (index = 0;
 6008        d->cur != d->end && (!sized_array_p || index <= max_index_cst);
 6009        ++index)
 6010     {
and update d.cur accordingly.  In other words, when reshape_init gets

{{col[0][0], col[1][0], col[2][0], col[3][0]},
 {col[0][1], col[1][1], col[2][1], col[3][1]},
 {col[0][2], col[1][2], col[2][2], col[3][2]},
 {col[0][3], col[1][3], col[2][3], col[3][3]}}

we recurse on the first element:
  {col[0][0], col[1][0], col[2][0], col[3][0]}
and we can't just move d.cur to point to
  {col[0][1], col[1][1], col[2][1], col[3][1]}
and return; we need to iterate, so that d.cur ends up being properly
updated, and after all initializers have been seen, points to d.end.
Currently we skip the loop, wherefore we hit this:

 6502   /* Make sure all the element of the constructor were used. Otherwise,
 6503      issue an error about exceeding initializers.  */
 6504   if (d.cur != d.end)
 6505     {
 6506       if (complain & tf_error)
 6507         error ("too many initializers for %qT", type);
 6508       return error_mark_node;
 6509     }

* decl.c (reshape_init_r): For a nested compound literal, do
call reshape_init_{class,array,vector}.

* g++.dg/cpp0x/initlist118.C: New test.

From-SVN: r279686

PR c++/92974 - bogus location for enum and non-enum in ?: warning.

build_min_non_dep wasn't setting any location so when we were emitting the
warning in the following test while instantiating a template, its location
was UNKNOWN_LOCATION.  Rather than adding a location_t parameter, let's use
the location from the original expression.

* tree.c (build_min_non_dep): Use the location of NON_DEP when
building the expression.

* g++.dg/diagnostic/enum1.C: New test.
* g++.dg/gomp/loop-2.C: Adjust dg-error.
* g++.dg/gomp/for-21.C: Likewise.

From-SVN: r279685

re PR c++/92965 ("note: 'x' is not public" emitted even when no error is emitted)

PR c++/92965
* pt.c (invalid_nontype_parm_type_p): Call structural_type_p with
explain=true only if emitting error.

* g++.dg/cpp2a/nontype-class27.C: New test.

From-SVN: r279684

re PR c++/92966 (Segfault on defaulted operator== with wrong return type)

PR c++/92966
* method.c (early_check_defaulted_comparison): Don't set
DECL_MAYBE_DELETED when returning false.

* g++.dg/cpp2a/spaceship-eq8.C: New test.

From-SVN: r279683

re PR c++/92973 (Silently accepting defaulted comparison operators in C++11 .. 17)

PR c++/92973
* method.c (early_check_defaulted_comparison): For C++17 and earlier
diagnose defaulted comparison operators.

* g++.dg/cpp0x/spaceship-eq1.C: New test.

From-SVN: r279682

re PR c++/92666 (bogus -Wunused-but-set-variable in gcov.c with -Wno-restrict)

PR c++/92666
* call.c (convert_arg_to_ellipsis): For floating point or
decltype(nullptr) arguments call mark_rvalue_use.

* g++.dg/warn/Wunused-var-36.C: New test.

From-SVN: r279681

re PR c++/92992 (Side-effects dropped when decltype(nullptr) typed expression is passed to ellipsis)

PR c++/92992
* call.c (convert_arg_to_ellipsis): For decltype(nullptr) arguments
that have side-effects use cp_build_compound_expr.

* g++.dg/cpp0x/nullptr45.C: New test.

From-SVN: r279680

Rename signed integer 16/34-bit macros.

2019-12-20   Michael Meissner  <meissner@linux.ibm.com>

* config/rs6000/predicates.md (cint34_operand): Use
SIGNED_INTEGER_34BIT_P macro.
* config/rs6000/rs6000.c (num_insns_constant_gpr): Use the
SIGNED_INTEGER_16BIT_P and SIGNED_INTEGER_34BIT_P macros.
(address_to_insn_form): Use the SIGNED_INTEGER_16BIT_P and
SIGNED_INTEGER_34BIT_P macros.
* config/rs6000/rs6000.h (SIGNED_INTEGER_NBIT_P): New macro.
(SIGNED_INTEGER_16BIT_P): Rename SIGNED_16BIT_OFFSET_P to be
SIGNED_INTEGER_34BIT_P.
(SIGNED_INTEGER_34BIT_P): Rename SIGNED_34BIT_OFFSET_P to be
SIGNED_INTEGER_34BIT_P.

From-SVN: r279677

c-ada-spec.h (decl_sloc): Delete.

c-family/
* c-ada-spec.h (decl_sloc): Delete.
* c-ada-spec.c (decl_sloc): Make static.
c/
* c-decl.c (collect_source_ref_cb): Delete.
(for_each_global_decl): Rename into...
(collect_source_refs): ...this.  Call collect_source_ref directly.
(c_parse_final_cleanups): Always call collect_source_ref on the main
input filename.
cp/
* decl2.c (c_parse_final_cleanups): Always call collect_source_ref on
the main input filename.

From-SVN: r279670

fortran: Fix PR number in comment of testcase for PR 69497

The testcase was originally committed with an incorrect changelog and PR
number. The changelog was fixed later, but not the comment in the test.

PR fortran/69497
* gfortran.dg/pr69497.f90: Fix PR number in comment.

From-SVN: r279657

libstdc++: Add inline to maybe-constexpr functions (PR 92927)

Originally these functions were always inline. I changed them in r277342
to be always constexpr, then in r277588 changed them to be constexpr for
C++14, but I didn't restore the 'inline' for C++11. That leads to linker
errors when libstdc++.so is built unoptimized, because those functions
don't get instantiated in src/c++11/string-inst.o

PR libstdc++/92927
* include/bits/alloc_traits.h (__alloc_on_copy, __alloc_on_move)
(__alloc_on_swap): Add inline specifier.

From-SVN: r279656

gcc/testsuite/ChangeLog:

2019-12-20  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* lib/target-supports.exp
(check_effective_target_arm_v8_2a_i8mm_ok_nocache): New.
(check_effective_target_arm_v8_2a_i8mm_ok): New.
(add_options_for_arm_v8_2a_i8mm): New.
(check_effective_target_arm_v8_2a_bf16_neon_ok_nocache): New.
(check_effective_target_arm_v8_2a_bf16_neon_ok): New.
(add_options_for_arm_v8_2a_bf16_neon): New.

gcc/Changelog:

2019-12-20  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* doc/sourcebuild.texi
(arm_v8_2a_bf16_neon_ok): Document new target supports option.
(arm_v8_2a_i8mm_ok): Likewise.

From-SVN: r279648

Rename tests to reference the correct PR

2019-12-20  Roman Zhuykov  <zhroma@ispras.ru>

* gcc.dg/pr92951-1.c: Rename to ...
* gcc.dg/pr92591-1.c: ... this.
* gcc.dg/pr92951-2.c: Rename to ...
* gcc.dg/pr92591-2.c: ... this.

From-SVN: r279645

libstdc++: Test setrlimit with c++ in configure

* acinclude.m4 (GLIBCXX_CHECK_SETRLIMIT): Test with AC_LANG_CPLUSPLUS.
* configure: Regenerate.

From-SVN: r279644

libstdc++: Fix versioned namespace tests

* testsuite/23_containers/map/48101_neg.cc: Add versioned namespace
pattern to tested error message.
* testsuite/23_containers/multimap/48101_neg.cc: Likewise.
* testsuite/30_threads/headers/stop_token/synopsis.cc: Add
dg-require-normal-namepace.

From-SVN: r279641

libstdc++: Fix pretty printers script and tests

* python/libstdcxx/v6/printers.py (lookup_node_type): Remove redundant
call to lookup_node_type.
* testsuite/libstdc++-prettyprinters/80276.cc: Define
_GLIBCXX_USE_CXX11_ABI to 0.
* testsuite/libstdc++-prettyprinters/91997.cc: Use regexp-test to check
'a' content.

From-SVN: r279640

Introduce -finline-arg-packing.

2019-12-20  Thomas Koenig  <tkoenig@gcc.gnu.org>

    PR middle-end/91512
    PR fortran/92738
    * invoke.texi: Document -finline-arg-packing.
    * lang.opt: Add -finline-arg-packing.
    * options.c (gfc_post_options): Handle -finline-arg-packing.
    * trans-array.c (gfc_conv_array_parameter): Use
    flag_inline_arg_packing instead of checking for optimize and
    optimize_size.

2019-12-20  Thomas Koenig  <tkoenig@gcc.gnu.org>

    PR middle-end/91512
    PR fortran/92738
    * gfortran.dg/inline_pack_25.f90: New test.

From-SVN: r279639

Fortran] PR 92996 – fix rank resolution EXPR_ARRAY

        PR fortran/92996
        gcc/fortran/
        * expr.c (simplify_parameter_variable): Call gfc_resolve_ref and
        gfc_expression_rank; fix location info.
        * gfortran.h (gfc_resolve_ref, gfc_expression_rank): Declare.
        * match.c (gfc_match_stopcode): Remove redundant setting of
        gfc_init_expr_flag; early return if gfc_simplify_expr has an error.
        * resolve.c (gfc_expression_rank): Renamed from expression_rank;
        minor cleanup.
        (gfc_resolve_ref): Removed static and renamed from resolve_ref.
        (resolve_variable, resolve_typebound_function,
        resolve_typebound_subroutine, resolve_ppc_call, resolve_expr_ppc,
        gfc_resolve_expr, resolve_procedure): Update calls.

        PR fortran/92996
        gcc/testsuite/
        * gfortran.dg/array_simplify_4.f90: New.
        * gfortran.dg/pr91565.f90: Update dg-error.
        * gfortran.dg/pr91801.f90: Likewise.

From-SVN: r279638

Improve is-coindexed check for OpenACC/OpenMP

        gcc/fortran/
        * openmp.c (resolve_omp_clauses): Move is-coindexed check from here ...
        (gfc_match_omp_variable_list): ... to here.

        gcc/testsuite/
        * gfortran.dg/goacc/coindexed-1.f90: New.

From-SVN: r279637

Add myself to MAINTAINERS

From-SVN: r279636

libstdc++: Strengthen the check for availability of pthread_rwlock_t

        * acinclude.m4 (_GLIBCXX_USE_PTHREAD_RWLOCK_T): Checks that
        _PTHREADS is defined after including gthr.h.
        * configure: Regenerate.

From-SVN: r279635

Fix testsuite-fallout of OpenACC deep-copy patch

        gcc/testsuite/
        * gfortran.dg/goacc/data-clauses.f95: Remove now
        obsolete dg-error.

From-SVN: r279634

re PR target/92841 (Optimize -fstack-protector-strong code generation a bit)

PR target/92841
* config/i386/i386.md (*stack_protect_set_3): For pic_32bit_operand
always use lea{q}, no matter what value which_alternative has.

* gcc.target/i386/pr92841-2.c: New test.

From-SVN: r279633

re PR target/93002 (while(i--) optimization)

PR target/93002
* config/i386/i386.md (dec reg; cmp $-1, reg; jne lab): New
define_peephole2.

* gcc.target/i386/pr93002.c: New test.

From-SVN: r279632

Fortran polymorphic class-type support for OpenACC

gcc/fortran/
* openmp.c (resolve_oacc_data_clauses): Don't disallow allocatable
polymorphic types for OpenACC.
* trans-openmp.c (gfc_trans_omp_clauses): Support polymorphic class
types.

libgomp/
* testsuite/libgomp.oacc-fortran/class-ptr-param.f95: New test.
* testsuite/libgomp.oacc-fortran/classtypes-1.f95: New test.
* testsuite/libgomp.oacc-fortran/classtypes-2.f95: New test.

From-SVN: r279631

OpenACC 2.6 deep copy: Fortran execution tests

libgomp/
* testsuite/libgomp.oacc-fortran/deep-copy-1.f90: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-2.f90: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-3.f90: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-4.f90: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-5.f90: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-6.f90: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-7.f90: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-8.f90: New test.
* testsuite/libgomp.oacc-fortran/derived-type-1.f90: New test.
* testsuite/libgomp.oacc-fortran/derivedtype-1.f95: New test.
* testsuite/libgomp.oacc-fortran/derivedtype-2.f95: New test.
* testsuite/libgomp.oacc-fortran/multidim-slice.f95: New test.
* testsuite/libgomp.oacc-fortran/update-2.f90: New test.

Co-Authored-By: Cesar Philippidis <cesar@codesourcery.com>
From-SVN: r279630

OpenACC 2.6 deep copy: C and C++ execution tests

libgomp/
* testsuite/libgomp.oacc-c-c++-common/deep-copy-1.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-2.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-4.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-6.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-7.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-8.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-9.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-10.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-11.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-14.c: New test.
* testsuite/libgomp.oacc-c++/deep-copy-12.C: New test.
* testsuite/libgomp.oacc-c++/deep-copy-13.C: New test.

From-SVN: r279629

OpenACC 2.6 deep copy: Fortran front-end parts

        gcc/fortran/
        * gfortran.h (gfc_omp_map_op): Add OMP_MAP_ATTACH, OMP_MAP_DETACH.
        * openmp.c (gfc_match_omp_variable_list): Add allow_derived parameter.
        Parse derived-type member accesses if true.
        (omp_mask2): Add OMP_CLAUSE_ATTACH and OMP_CLAUSE_DETACH.
        (gfc_match_omp_map_clause): Add allow_derived parameter.  Pass to
        gfc_match_omp_variable_list.
        (gfc_match_omp_clauses): Support attach and detach.  Support derived
        types for appropriate OpenACC directives.
        (OACC_PARALLEL_CLAUSES, OACC_SERIAL_CLAUSES, OACC_KERNELS_CLAUSES,
        OACC_DATA_CLAUSES, OACC_ENTER_DATA_CLAUSES): Add OMP_CLAUSE_ATTACH.
        (OACC_EXIT_DATA_CLAUSES): Add OMP_CLAUSE_DETACH.
        (check_symbol_not_pointer): Don't disallow pointer objects of derived
        type.
        (resolve_oacc_data_clauses): Don't disallow allocatable derived types.
        (resolve_omp_clauses): Perform duplicate checking only for non-derived
        type component accesses (plain variables and arrays or array sections).
        Support component refs.
        * trans-expr.c (gfc_conv_component_ref,
        conv_parent_component_references): Make global.
        (gfc_maybe_dereference_var): New function, broken out of...
        (gfc_conv_variable): ...here.  Call above function.
        * trans-openmp.c (gfc_omp_privatize_by_reference): Support component
        refs.
        (gfc_trans_omp_array_section): New function, broken out of...
        (gfc_trans_omp_clauses): ...here.  Support component refs/derived
        types, attach and detach clauses.
        * trans.h (gfc_conv_component_ref, conv_parent_component_references,
        gfc_maybe_dereference_var): Add prototypes.

        gcc/testsuite/
        * gfortran.dg/goacc/derived-types.f90: New test.
        * gfortran.dg/goacc/derived-types-2.f90: New test.
        * gfortran.dg/goacc/derived-types-3.f90: New test.
        * gfortran.dg/goacc/data-clauses.f95: Adjust for expected errors.
        * gfortran.dg/goacc/enter-exit-data.f95: Likewise.

From-SVN: r279628

OpenACC 2.6 deep copy: C and C++ front-end parts

gcc/c-family/
* c-common.h (c_omp_map_clause_name): Add prototype.
* c-omp.c (c_omp_map_clause_name): New function.
* c-pragma.h (pragma_omp_clause): Add PRAGMA_OACC_CLAUSE_ATTACH and
PRAGMA_OACC_CLAUSE_DETACH.

gcc/c/
* c-parser.c (c_parser_omp_clause_name): Add parsing of attach and
detach clauses.
(c_parser_omp_variable_list): Add ALLOW_DEREF optional parameter.
Allow deref (->) in variable lists if true.
(c_parser_omp_var_list_parens): Add ALLOW_DEREF optional parameter.
Pass to c_parser_omp_variable_list.
(c_parser_oacc_data_clause): Support attach and detach clauses.  Update
call to c_parser_omp_variable_list.
(c_parser_oacc_all_clauses): Support attach and detach clauses.
(OACC_DATA_CLAUSE_MASK, OACC_ENTER_DATA_CLAUSE_MASK,
OACC_KERNELS_CLAUSE_MASK, OACC_PARALLEL_CLAUSE_MASK,
OACC_SERIAL_CLAUSE_MASK): Add PRAGMA_OACC_CLAUSE_ATTACH.
(OACC_EXIT_DATA_CLAUSE_MASK): Add PRAGMA_OACC_CLAUSE_DETACH.
* c-typeck.c (handle_omp_array_sections_1): Reject subarrays for attach
and detach.  Support deref.
(handle_omp_array_sections): Use GOMP_MAP_ATTACH_DETACH instead of
GOMP_MAP_ALWAYS_POINTER for OpenACC.
(c_oacc_check_attachments): New function.
(c_finish_omp_clauses): Check attach/detach arguments for being
pointers using above.  Support deref.

gcc/cp/
* parser.c (cp_parser_omp_clause_name): Support attach and detach
clauses.
(cp_parser_omp_var_list_no_open): Add ALLOW_DEREF optional parameter.
Parse deref if true.
(cp_parser_omp_var_list): Add ALLOW_DEREF optional parameter.  Pass to
cp_parser_omp_var_list_no_open.
(cp_parser_oacc_data_clause): Support attach and detach clauses.
Update call to cp_parser_omp_var_list_no_open.
(cp_parser_oacc_all_clauses): Support attach and detach.
(OACC_DATA_CLAUSE_MASK, OACC_ENTER_DATA_CLAUSE_MASK,
OACC_KERNELS_CLAUSE_MASK, OACC_PARALLEL_CLAUSE_MASK,
OACC_SERIAL_CLAUSE_MASK): Add PRAGMA_OACC_CLAUSE_ATTACH.
(OACC_EXIT_DATA_CLAUSE_MASK): Add PRAGMA_OACC_CLAUSE_DETACH.
* semantics.c (handle_omp_array_sections_1): Reject subarrays for
attach and detach.
(handle_omp_array_sections): Use GOMP_MAP_ATTACH_DETACH instead of
GOMP_MAP_ALWAYS_POINTER for OpenACC.
(cp_oacc_check_attachments): New function.
(finish_omp_clauses): Use above function.  Allow structure fields and
class members to appear in OpenACC data clauses.  Support
GOMP_MAP_ATTACH_DETACH.  Support deref.

gcc/testsuite/
* c-c++-common/goacc/deep-copy-arrayofstruct.c: New test.
* c-c++-common/goacc/mdc-1.c: New test.
* c-c++-common/goacc/mdc-2.c: New test.
* gcc.dg/goacc/mdc.C: New test.

Co-Authored-By: Cesar Philippidis <cesar@codesourcery.com>
From-SVN: r279627

OpenACC 2.6 deep copy: middle-end parts

gcc/
* gimplify.c (gimplify_omp_var_data): Add GOVD_MAP_HAS_ATTACHMENTS.
(insert_struct_comp_map): Support derived-type member mappings
for arrays with descriptors which use GOMP_MAP_TO_PSET.  Support
GOMP_MAP_ATTACH_DETACH.
(gimplify_scan_omp_clauses): Tidy up OACC_ENTER_DATA/OACC_EXIT_DATA
mappings.  Handle attach/detach clauses and component references.
(gimplify_adjust_omp_clauses_1): Skip adjustments for explicit
attach/detach clauses.
(gimplify_omp_target_update): Handle struct mappings and finalize for
detach operations.
* omp-low.c (lower_omp_target): Support GOMP_MAP_ATTACH,
GOMP_MAP_DETACH, GOMP_MAP_FORCE_DETACH.
* tree-pretty-print.c (dump_omp_clause): Likewise, plus
GOMP_MAP_ATTACH_DETACH.

include/
* gomp-constants.h (gomp_map_kind): Add GOMP_MAP_ATTACH_DETACH.

From-SVN: r279626

OpenACC 2.6 deep copy: libgomp parts

include/
* gomp-constants.h (GOMP_MAP_FLAG_SPECIAL_4, GOMP_MAP_DEEP_COPY):
Define.
(gomp_map_kind): Add GOMP_MAP_ATTACH, GOMP_MAP_DETACH,
GOMP_MAP_FORCE_DETACH.

libgomp/
* libgomp.h (struct target_var_desc): Add do_detach flag.
* oacc-init.c (acc_shutdown_1): Free aux block if present.
* oacc-mem.c (find_group_last): Add SIZES parameter. Support
struct components.  Tidy up and add some new checks.
(goacc_enter_data_internal): Update call to find_group_last.
(goacc_exit_data_internal): Support detach operations and
GOMP_MAP_STRUCT.
(GOACC_enter_exit_data): Handle initial GOMP_MAP_STRUCT or
GOMP_MAP_FORCE_PRESENT in finalization detection code.  Handle
attach/detach in enter/exit data detection code.
* target.c (gomp_map_vars_existing): Initialise do_detach field of
tgt_var_desc.
(gomp_map_vars_internal): Support attach.
(gomp_unmap_vars_internal): Support detach.

From-SVN: r279625

OpenACC 2.6 deep copy: attach/detach API routines

libgomp/
* libgomp.h (struct splay_tree_aux): Add attach_count field.
(gomp_attach_pointer, gomp_detach_pointer): Add prototypes.
* libgomp.map (OACC_2.6): New section. Add acc_attach,
acc_attach_async, acc_detach, acc_detach_async, acc_detach_finalize,
acc_detach_finalize_async.
* oacc-mem.c (acc_attach_async, acc_attach, goacc_detach_internal,
acc_detach, acc_detach_async, acc_detach_finalize,
acc_detach_finalize_async): New functions.
* openacc.h (acc_attach, acc_attach_async, acc_detach,
(acc_detach_async, acc_detach_finalize, acc_detach_finalize_async): Add
prototypes.
* target.c (gomp_attach_pointer, gomp_detach_pointer): New functions.
(gomp_remove_var_internal): Free attachment counts if present.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-3.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-5.c: New test.

Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
From-SVN: r279624

Factor out duplicate code in gimplify_scan_omp_clauses

gcc/
* gimplify.c (insert_struct_comp_map, extract_base_bit_offset): New.
(gimplify_scan_omp_clauses): Outline duplicated code into calls to
above two functions.

From-SVN: r279623

Use gomp_map_val for OpenACC host-to-device address translation

libgomp/
* libgomp.h (gomp_map_val): Add prototype.
* oacc-parallel.c (GOACC_parallel_keyed): Use gomp_map_val instead of
open-coding device-address calculation.
* target.c (gomp_map_val): Make global. Use OFFSET_POINTER in
non-present case.

Co-Authored-By: Cesar Philippidis <cesar@codesourcery.com>
From-SVN: r279622

OpenACC reference count overhaul

libgomp/
* libgomp.h (struct splay_tree_key_s): Substitute dynamic_refcount
field for virtual_refcount.
(enum gomp_map_vars_kind): Add GOMP_MAP_VARS_OPENACC_ENTER_DATA.
(gomp_free_memmap): Remove prototype.
* oacc-init.c (acc_shutdown_1): Iteratively call gomp_remove_var
instead of calling gomp_free_memmap.
* oacc-mem.c (acc_map_data): Use virtual_refcount instead of
dynamic_refcount.
(acc_unmap_data): Open code instead of forcing target_mem_desc's
to_free field to NULL then calling gomp_unmap_vars.  Handle
REFCOUNT_INFINITY on target blocks.
(goacc_enter_data): Rename to...
(goacc_enter_datum): ...this.  Remove MAPNUM parameter and special
handling for mapping groups.  Use virtual_refcount instead of
dynamic_refcount.  Use GOMP_MAP_VARS_OPENACC_ENTER_DATA for
map_map_vars_async call.  Re-do lookup for target pointer return value.
(acc_create, acc_create_async, acc_copyin, acc_copyin_async): Call
renamed goacc_enter_datum function.
(goacc_exit_data): Rename to...
(goacc_exit_datum): ...this.  Update for virtual_refcount semantics.
(acc_delete, acc_delete_async, acc_delete_finalize,
acc_delete_finalize_async, acc_copyout, acc_copyout_async,
acc_copyout_finalize, acc_copyout_finalize_async): Call renamed
goacc_exit_datum function.
(gomp_acc_remove_pointer, find_pointer): Remove functions.
(find_group_last, goacc_enter_data_internal, goacc_exit_data_internal):
New functions.
(GOACC_enter_exit_data): Use goacc_enter_data_internal and
goacc_exit_data_internal helper functions.
* target.c (gomp_map_vars_internal): Handle
GOMP_MAP_VARS_OPENACC_ENTER_DATA.  Update for virtual_refcount
semantics.
(gomp_unmap_vars_internal): Update for virtual_refcount semantics.
(gomp_load_image_to_device, omp_target_associate_ptr): Zero-initialise
virtual_refcount field instead of dynamic_refcount.
(gomp_free_memmap): Remove function.
* testsuite/libgomp.oacc-c-c++-common/unmap-infinity-1.c: New test.
* testsuite/libgomp.c-c++-common/unmap-infinity-2.c: New test.
* testsuite/libgomp.oacc-c-c++-common/pr92843-1.c: Add XFAIL.

From-SVN: r279621

Use aux struct in libgomp for infrequently-used/API-specific data

libgomp/
* libgomp.h (struct splay_tree_aux): New.
(struct splay_tree_key_s): Replace link_key field with aux pointer.
* target.c (gomp_map_vars_internal): Adjust for link_key being moved
to aux struct.
(gomp_remove_var_internal): Free aux block if present.
(gomp_load_image_to_device): Zero-initialise aux field instead of
link_key field.
(omp_target_associate_pointer): Zero-initialise aux field.

Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
From-SVN: r279620

Daily bump.

From-SVN: r279619

re PR target/92905 (Spills float-int union to memory)

2019-12-19  Vladimir Makarov  <vmakarov@redhat.com>

PR target/92905
* lra-constraints.c (process_alt_operands): Check offmemok when
processing preferred_reload_class.

2019-12-19  Vladimir Makarov  <vmakarov@redhat.com>

PR target/92905
* gcc.target/i386/pr92905.c: New test.

From-SVN: r279596

Update g++.dg/ext/sve-sizeless-2.C after r279471

2019-12-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/testsuite/
* g++.dg/ext/sve-sizeless-2.C: Don't expect an error for
alias templates.

From-SVN: r279588

Allow constants in amdgcn extends and truncates

2019-12-19  Andrew Stubbs  <ams@codesourcery.com>

gcc/
* config/gcn/gcn-valu.md
(<convop><VEC_ALL1REG_INT_ALT:mode><VEC_ALL1REG_INT_MODE:mode>2<exec>):
Change input predcate to gcn_alu_operand.
(extend<VEC_ALL1REG_INT_ALT:mode><VEC_ALL1REG_INT_MODE:mode>2<exec>):
Likewise.
(truncv64di<mode>2): Likewise.
(truncv64di<mode>2_exec): Likewise.
(<convop><mode>v64di2): Likewise.
(<convop><mode>v64di2_exec): Likewise.

From-SVN: r279587

Use V64SI for all amdgcn add-with-carry insns

2019-12-19  Andrew Stubbs  <ams@codesourcery.com>

gcc/
* config/gcn/gcn-valu.md (*plus_carry_dpp_shr_<mode>): Rename to ...
(*plus_carry_dpp_shr_v64si): ... this, and replace all
VEC_1REG_INT_MODE with V64SI.

From-SVN: r279584

Prevent conversion of character data in array constructors.

Fix for PR fortran/92896 [10 Regression] [DEC] ICE in reduce_unary, at
fortran/arith.c:1283.

This was caused by an unintended side affect of "Allow CHARACTER literals
in assignments and data statements" (revision 277975). If the conversion
occurs in a array constructor it is rejected.

From-SVN: r279583

hash-map-tests.c: add a selftest involving int_hash

gcc/ChangeLog:
* hash-map-tests.c (selftest::test_map_of_int_to_strings): New
selftest.
(selftest::hash_map_tests_c_tests): Call it.

From-SVN: r279582

gimple const-correctness fixes

This patch converts various "gimple *" to "const gimple *" and similar
fixes for gimple subclasses, adding is_a_helper for gimple subclasses
to support the const form of as_a, and adding a few "const" overloads
of accessors.

This is enough to make pp_gimple_stmt_1's stmt const.

gcc/ChangeLog:
* gimple-predict.h (gimple_predict_predictor): Make "gs" param
const.
(gimple_predict_outcome): Likewise.
* gimple-pretty-print.c (do_niy): Likewise.
(dump_unary_rhs): Likewise.
(dump_binary_rhs): Likewise.
(dump_ternary_rhs): Likewise.
(dump_gimple_assign): Likewise.
(dump_gimple_return): Likewise.
(dump_gimple_call_args): Likewise.
(pp_points_to_solution): Make "pt" param const.
(dump_gimple_call): Make "gs" param const.
(dump_gimple_switch): Likewise.
(dump_gimple_cond): Likewise.
(dump_gimple_label): Likewise.
(dump_gimple_goto): Likewise.
(dump_gimple_bind): Likewise.
(dump_gimple_try): Likewise.
(dump_gimple_catch): Likewise.
(dump_gimple_eh_filter): Likewise.
(dump_gimple_eh_must_not_throw): Likewise.
(dump_gimple_eh_else): Likewise.
(dump_gimple_resx): Likewise.
(dump_gimple_eh_dispatch): Likewise.
(dump_gimple_debug): Likewise.
(dump_gimple_omp_for): Likewise.
(dump_gimple_omp_continue): Likewise.
(dump_gimple_omp_single): Likewise.
(dump_gimple_omp_taskgroup): Likewise.
(dump_gimple_omp_target): Likewise.
(dump_gimple_omp_teams): Likewise.
(dump_gimple_omp_sections): Likewise.
(dump_gimple_omp_block): Likewise.
(dump_gimple_omp_critical): Likewise.
(dump_gimple_omp_ordered): Likewise.
(dump_gimple_omp_scan): Likewise.
(dump_gimple_omp_return): Likewise.
(dump_gimple_transaction): Likewise.
(dump_gimple_asm): Likewise.
(dump_gimple_phi): Make "phi" param const.
(dump_gimple_omp_parallel): Make "gs" param const.
(dump_gimple_omp_task): Likewise.
(dump_gimple_omp_atomic_load): Likewise.
(dump_gimple_omp_atomic_store): Likewise.
(dump_gimple_mem_ops): Likewise.
(pp_gimple_stmt_1): Likewise.  Add "const" to the various as_a <>
casts throughout.
* gimple-pretty-print.h (gimple_stmt_1): Make gimple * param const.
* gimple.h (is_a_helper <const gdebug *>::test): New.
(is_a_helper <const ggoto *>::test): New.
(is_a_helper <const glabel *>::test): New.
(is_a_helper <const geh_else *>::test): New.
(is_a_helper <const geh_mnt *>::test): New.
(is_a_helper <const gswitch *>::test): New.
(is_a_helper <const gtry *>::test): New.
(is_a_helper <const greturn *>::test): New.
(gimple_call_tail_p): Make param const.
(gimple_call_return_slot_opt_p): Likewise.
(gimple_call_va_arg_pack_p): Likewise.
(gimple_call_use_set): Add const overload.
(gimple_call_clobber_set): Likewise.
(gimple_has_lhs): Make param const.
(gimple_bind_body): Likewise.
(gimple_catch_handler): Likewise.
(gimple_eh_filter_failure): Likewise.
(gimple_eh_must_not_throw_fndecl): Likewise.
(gimple_eh_else_n_body): Likewise.
(gimple_eh_else_e_body): Likewise.
(gimple_try_eval): Likewise.
(gimple_try_cleanup): Likewise.
(gimple_phi_arg): Add const overload.
(gimple_phi_arg_def): Make param const.
(gimple_phi_arg_edge): Likewise.
(gimple_phi_arg_location): Likewise.
(gimple_phi_arg_has_location): Likewise.
(gimple_debug_bind_get_var): Likewise.
(gimple_debug_bind_get_value): Likewise.
(gimple_debug_source_bind_get_var): Likewise.
(gimple_debug_source_bind_get_value): Likewise.
(gimple_omp_body): Likewise.
(gimple_omp_for_collapse): Likewise.
(gimple_omp_for_pre_body): Likewise.
(gimple_transaction_body): Likewise.
* tree-eh.c (lookup_stmt_eh_lp_fn): Make param "t" const.
(lookup_stmt_eh_lp): Likewise.
* tree-eh.h (lookup_stmt_eh_lp_fn): Make param const.
(lookup_stmt_eh_lp): Likewise.
* tree-ssa-alias.h (pt_solution_empty_p): Make param const.
* tree-ssa-structalias.c (pt_solution_empty_p): Likewise.

From-SVN: r279581

[AArch64] Fix handling of npatterns>1 constants for partial SVE modes

For partial SVE vectors of element X, we want to treat duplicates
of single X elements in the same way as for full vectors of X.
But if a constant instead contains a repeating pattern of X elements,
the transition from one value to the next must happen at container
boundaries rather than element boundaries.  E.g. a VNx4HI should
in that case contain the same number of constants as a VNx4SI.

Fixing this means that we need a reinterpret from the container-based
mode to the partial mode; e.g. in the above example we need a
reinterpret from VNx4SI to VNx4HI.  We can't use subregs for that
because they're forbidden by aarch64_can_change_class_mode; we should
handle them in the same way as for big-endian instead.

2019-12-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.c (aarch64_simd_valid_immediate): When
handling partial SVE vectors, use the container mode rather than
the element mode if the constant isn't a single-element duplicate.
* config/aarch64/aarch64-sve.md (@aarch64_sve_reinterpret<mode>):
Check targetm.can_change_mode_class instead of BYTES_BIG_ENDIAN.

gcc/testsuite/
* gcc.target/aarch64/sve/mixed_size_9.c: New test.

From-SVN: r279580

comment tweak

From-SVN: r279579

PR c++/52320 - EH cleanups for partially constructed arrays.

This testcase wasn't fixed by the 66139 patch; split_nonconstant_init_1 was
failing to add a cleanup for an array member of a class (e.g. e1) that will
run if initializing a later member (e.g. e2) throws.

* typeck2.c (split_nonconstant_init_1): Add nested parm.
Add cleanup for whole array if true.

From-SVN: r279577

PR c++/66139 - EH cleanups for partially constructed aggregates.

There were several overlapping PRs about failure to clean up fully
constructed subobjects when an exception is thrown during aggregate
initialization of a temporary.  I fixed this for non-temporaries in the
context of 57510, but that fix didn't handle temporaries.  So this patch
does split_nonconstant_init at gimplification time, which is much smaller
than alternatives I tried.

PR c++/57510
* cp-gimplify.c (cp_gimplify_init_expr): Use split_nonconstant_init.
* typeck2.c (split_nonconstant_init): Handle non-variable dest.
(split_nonconstant_init_1): Clear TREE_SIDE_EFFECTS.
* tree.c (is_local_temp): New.

From-SVN: r279576

Add changelog for commit r279573 (added myself to write-after-approval)

2019-12-19  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* MAINTAINERS (write_after_approval): Add myself.

From-SVN: r279575

Implement sub-dword add/sub on amdgcn

2019-12-19  Andrew Stubbs  <ams@codesourcery.com>

gcc/
* config/gcn/gcn-valu.md (addv64si3<exec_clobber>): Rename to ...
(add<mode>3<exec_clobber>): ... this, and use VEC_ALL1REG_INT_MODE.
(addv64si3_dup<exec_clobber>): Rename to ...
(add<mode>3_dup<exec_clobber>): ... this, and use VEC_ALL1REG_INT_MODE.
(subv64si3<exec_clobber>): Rename to ...
(sub<mode>3<exec_clobber>): ... this, and use VEC_ALL1REG_INT_MODE.

From-SVN: r279574

Add myself to MAINTAINERS

From-SVN: r279573

[AArch64] Reject invalid subregs involving partial SVE modes

When adding partial SVE modes, I'd remembered to handle reloads
in a similar way to full big-endian SVE vectors, but forgot the
just-as-important mode-change rules.

2019-12-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.c (aarch64_can_change_mode_class):
Don't allow changes between partial SVE modes and other kinds
of mode.  Don't allow changes between two partial SVE modes
if they have different container or element sizes.

gcc/testsuite/
* gcc.target/aarch64/sve/mixed_size_8.c: New test.

From-SVN: r279572

[AArch64] Handle arguments and return types with partial SVE modes

Partial SVE modes can be picked up and used by the vector_size(N)
attribute.[*] This means that we need to cope with arguments and return
values with partial SVE modes, which previously triggered asserts like:

  /* Generic vectors that map to SVE modes with -msve-vector-bits=N are
     passed by reference, not by value.  */
  gcc_assert (!aarch64_sve_mode_p (mode));

The ABI for these types is fixed from pre-SVE days, and must in any case
be the same for all -msve-vector-bits=N values.  All we need to do is
ensure that the vectors are passed and returned in the traditional way.

[*] Advanced SIMD always wins for 64-bit and 128-bit vectors though.

2019-12-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.c (aarch64_function_value_1): New function,
split out from...
(aarch64_function_value): ...here.  Handle partial SVE modes by
pretending that they have the associated/traditional integer mode,
then wrap the result in the real mode.
(aarch64_layout_arg): Take an orig_mode argument and pass it to
aarch64_function_arg_alignment.  Handle partial SVE modes analogously
to aarch64_function_value.
(aarch64_function_arg): Update call accordingly.
(aarch64_function_arg_advance): Likewise.

gcc/testsuite/
* gcc.target/aarch64/sve/pcs/gnu_vectors_3.c: New test.

From-SVN: r279571

Revert accidental commit

From-SVN: r279570

Don't mangle attributes that have a space in their name

The SVE port needs to maintain a different type identity for
GNU vectors and "SVE vectors" even during LTO, since the types
use different ABIs.  The easiest way of doing that seemed to be
to use type attributes.  However, these type attributes shouldn't
be user-facing; they're just a convenient way of representing the
types internally in GCC.

There are already several internal-only attributes, such as "fn spec"
and "omp declare simd".  They're distinguished from normal user-facing
attributes by having a space in their name, which means that it isn't
possible to write them directly in C or C++.

Taking the same approach mostly works well for SVE.  The only snag
I've hit so far is that the new attribute needs to (and only exists to)
affect type identity.  This means that it would normally get included
in mangled names, to distinguish it from types without the attribute.

However, the SVE ABI specifies a separate mangling for SVE vector types,
rather than using an attribute mangling + a normal vector mangling.
So we need some way of suppressing the attribute mangling for this case.

There are currently no other target-independent or target-specific
internal-only attributes that affect type identity, so this patch goes
for the simplest fix of skipping mangling for attributes whose names
contain a space.  Other options I thought about were:

(1) Also make sure that targetm.mangled_type returns nonnull.

(2) Check directly for the target-specific name.

(3) Add a new target hook.

(4) Add new information to attribute_spec.  This would be very invasive
    at this stage, but maybe we should consider replacing all the boolean
    fields with flags?  That should make the tables slightly easier to
    read and would make adding new flags much simpler in future.

2019-12-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/cp/
* mangle.c (write_CV_qualifiers_for_type): Don't mangle attributes
that contain a space.

From-SVN: r279569

re PR fortran/92977 (ICE in gfc_trans_omp_atomic, at fortran/trans-openmp.c:3526)

PR fortran/92977
* frontend-passes.c (call_external_blas): Use || instead of |.

From-SVN: r279567

Fix symver attribute with LTO

* cgraph.c (cgraph_node_cannot_be_local_p_1): Prevent targets of
symver attributes to be localized.
* ipa-visibility.c (cgraph_externally_visible_p,
varpool_node::externally_visible_p): Likewise.
* symtab.c (symtab_node::verify_base): Check visibility of symbol
versions.

* lto-common.c (read_cgraph_and_symbols): Work around binutils
PR25424

Co-Authored-By: Xi Ruoyao <xry111@mengyan1223.wang>
From-SVN: r279566

ipa: duplicate ipa_size_summary for cloned nodes

* ipa-fnsummary.h (ipa_size_summary): Remove copy consturctor.
(ipa_size_summary_t): Add duplicate method; move to heap.
* ipa-fnsumary.c (ipa_fn_summary_alloc): Fix allocation.

From-SVN: r279563

Handle aggregate pass-through for self-recursive call (PR ipa/92794)

2019-12-19  Feng Xue  <fxue@os.amperecomputing.com>

        PR ipa/92794
        * ipa-cp.c (self_recursive_agg_pass_through_p): New function.
        (intersect_with_plats): Use error_mark_node as place holder
        when aggregate jump function is simple pass-through for
        self-recursive call.
        (intersect_with_agg_replacements): Likewise.
        (intersect_aggregates_with_edge): Likewise.
        (find_aggregate_values_for_callers_subset): Likewise.

2019-12-19  Feng Xue  <fxue@os.amperecomputing.com>

        PR ipa/92794
        * gcc.dg/ipa/92794.c: New test.

From-SVN: r279561

Daily bump.

From-SVN: r279560

PR c++/91165 follow-on tweak

I talked in the PR about possibly stripping the location from the args in
the hash table, since if we use the cache the locations would be wrong, but
didn't actually do anything about that.  Then I noticed that there's already
unshare_expr_without_location...

* constexpr.c (cxx_eval_call_expression): Use
unshare_expr_without_location.

From-SVN: r279557