Optimize memory broadcast for constant vector under AVX512.

For constant vector having one duplicated value, there's no need to put
whole vector in the constant pool, using embedded broadcast instead.

2020-07-09  Hongtao Liu  <hongtao.liu@intel.com>

gcc/ChangeLog:

PR target/87767
* config/i386/i386-features.c
(replace_constant_pool_with_broadcast): New function.
(constant_pool_broadcast): Ditto.
(class pass_constant_pool_broadcast): New pass.
(make_pass_constant_pool_broadcast): Ditto.
(remove_partial_avx_dependency): Call
replace_constant_pool_with_broadcast under TARGET_AVX512F, it
would save compile time when both pass rpad and cpb are
available.
(remove_partial_avx_dependency_gate): New function.
(class pass_remove_partial_avx_dependency::gate): Call
remove_partial_avx_dependency_gate.
* config/i386/i386-passes.def: Insert new pass after combine.
* config/i386/i386-protos.h
(make_pass_constant_pool_broadcast): Declare.
* config/i386/sse.md (*avx512dq_mul<mode>3<mask_name>_bcst):
New define_insn.
(*avx512f_mul<mode>3<mask_name>_bcst): Ditto.
* config/i386/avx512fintrin.h (_mm512_set1_ps,
_mm512_set1_pd,_mm512_set1_epi32, _mm512_set1_epi64): Adjusted.

gcc/testsuite/ChangeLog:

PR target/87767
* gcc.target/i386/avx2-broadcast-pr87767-1.c: New test.
* gcc.target/i386/avx512f-broadcast-pr87767-1.c: New test.
* gcc.target/i386/avx512f-broadcast-pr87767-2.c: New test.
* gcc.target/i386/avx512f-broadcast-pr87767-3.c: New test.
* gcc.target/i386/avx512f-broadcast-pr87767-4.c: New test.
* gcc.target/i386/avx512f-broadcast-pr87767-5.c: New test.
* gcc.target/i386/avx512f-broadcast-pr87767-6.c: New test.
* gcc.target/i386/avx512f-broadcast-pr87767-7.c: New test.
* gcc.target/i386/avx512vl-broadcast-pr87767-1.c: New test.
* gcc.target/i386/avx512vl-broadcast-pr87767-1.c: New test.
* gcc.target/i386/avx512vl-broadcast-pr87767-2.c: New test.
* gcc.target/i386/avx512vl-broadcast-pr87767-3.c: New test.
* gcc.target/i386/avx512vl-broadcast-pr87767-4.c: New test.
* gcc.target/i386/avx512vl-broadcast-pr87767-5.c: New test.
* gcc.target/i386/avx512vl-broadcast-pr87767-6.c: New test.

Adjust testcase.

gcc/testsuite/ChangeLog:
PR target/96246
PR target/96855
PR target/96856
PR target/96857
* g++.target/i386/avx512bw-pr96246-2.C: Add runtime check for
AVX512BW.
* g++.target/i386/avx512vl-pr96246-2.C: Add runtime check for
AVX512BW and AVX512VL
* g++.target/i386/avx512f-helper.h: New header.
* gcc.target/i386/pr92658-avx512f.c: Add
-mprefer-vector-width=512 to avoid impact of different default
mtune which gcc is built with.
* gcc.target/i386/avx512bw-pr95488-1.c: Ditto.
* gcc.target/i386/pr92645-4.c: Add -mno-avx512f to avoid
impact of different default march which gcc is built with.

Daily bump.

d: __vectors unsupported in hardware should be rejected at compile-time.

gcc/d/ChangeLog:

PR d/96869
* d-builtins.cc (build_frontend_type): Don't expose intrinsics that
use unsupported vector types.
* d-target.cc (Target::isVectorTypeSupported): Restrict to supporting
only if TARGET_VECTOR_MODE_SUPPORTED_P is true.  Don't allow complex
or boolean vector types.

gcc/testsuite/ChangeLog:

PR d/96869
* gdc.dg/simd.d: Removed.
* gdc.dg/cast1.d: New test.
* gdc.dg/gdc213.d: Compile with target vect_sizes_16B_8B.
* gdc.dg/gdc284.d: Likewise.
* gdc.dg/gdc67.d: Likewise.
* gdc.dg/pr96869.d: New test.
* gdc.dg/simd1.d: New test.
* gdc.dg/simd10447.d: New test.
* gdc.dg/simd12776.d: New test.
* gdc.dg/simd13841.d: New test.
* gdc.dg/simd13927.d: New test.
* gdc.dg/simd15123.d: New test.
* gdc.dg/simd15144.d: New test.
* gdc.dg/simd16087.d: New test.
* gdc.dg/simd16697.d: New test.
* gdc.dg/simd17237.d: New test.
* gdc.dg/simd17695.d: New test.
* gdc.dg/simd17720a.d: New test.
* gdc.dg/simd17720b.d: New test.
* gdc.dg/simd19224.d: New test.
* gdc.dg/simd19627.d: New test.
* gdc.dg/simd19628.d: New test.
* gdc.dg/simd19629.d: New test.
* gdc.dg/simd19630.d: New test.
* gdc.dg/simd2a.d: New test.
* gdc.dg/simd2b.d: New test.
* gdc.dg/simd2c.d: New test.
* gdc.dg/simd2d.d: New test.
* gdc.dg/simd2e.d: New test.
* gdc.dg/simd2f.d: New test.
* gdc.dg/simd2g.d: New test.
* gdc.dg/simd2h.d: New test.
* gdc.dg/simd2i.d: New test.
* gdc.dg/simd2j.d: New test.
* gdc.dg/simd7951.d: New test.
* gdc.dg/torture/array2.d: New test.
* gdc.dg/torture/array3.d: New test.
* gdc.dg/torture/simd16488a.d: New test.
* gdc.dg/torture/simd16488b.d: New test.
* gdc.dg/torture/simd16703.d: New test.
* gdc.dg/torture/simd19223.d: New test.
* gdc.dg/torture/simd19607.d: New test.
* gdc.dg/torture/simd3.d: New test.
* gdc.dg/torture/simd4.d: New test.
* gdc.dg/torture/simd7411.d: New test.
* gdc.dg/torture/simd7413a.d: New test.
* gdc.dg/torture/simd7413b.d: New test.
* gdc.dg/torture/simd7414.d: New test.
* gdc.dg/torture/simd9200.d: New test.
* gdc.dg/torture/simd9304.d: New test.
* gdc.dg/torture/simd9449.d: New test.
* gdc.dg/torture/simd9910.d: New test.

d: Only test with default permutation flags for runnable tests.

Unless the test explicitly requests, all compilable tests as well as
fail_compilation tests will be ran without any extra flags.

The C++ tests now are checked against shared D runtime library.

gcc/testsuite/ChangeLog:

* lib/gdc-utils.exp (gdc-convert-test): Handle LINK directive.
Set PERMUTE_ARGS as DEFAULT_DFLAGS only for runnable tests.
(gdc-do-test): Set default action of compilable tests to compile.
Test SHARED_OPTION on runnable_cxx tests.

d: Move all runnable tests in gdc.dg to gdc.dg/torture

Tests that are not executed do not need to be compiled as torture tests,
they are only present for testing for a certain bug or ICE.

gcc/testsuite/ChangeLog:

* gdc.dg/dg.exp: Remove torture options.
* gdc.dg/gdc115.d: Move test to gdc.dg/torture.
* gdc.dg/gdc131.d: Likewise.
* gdc.dg/gdc141.d: Likewise.
* gdc.dg/gdc17.d: Likewise.
* gdc.dg/gdc171.d: Likewise.
* gdc.dg/gdc179.d: Likewise.
* gdc.dg/gdc186.d: Likewise.
* gdc.dg/gdc187.d: Likewise.
* gdc.dg/gdc191.d: Likewise.
* gdc.dg/gdc198.d: Likewise.
* gdc.dg/gdc200.d: Likewise.
* gdc.dg/gdc210.d: Likewise.
* gdc.dg/gdc240.d: Likewise.
* gdc.dg/gdc242b.d: Likewise.
* gdc.dg/gdc248.d: Likewise.
* gdc.dg/gdc250.d: Likewise.
* gdc.dg/gdc273.d: Likewise.
* gdc.dg/gdc283.d: Likewise.
* gdc.dg/gdc285.d: Likewise.
* gdc.dg/gdc286.d: Likewise.
* gdc.dg/gdc309.d: Likewise.
* gdc.dg/gdc35.d: Likewise.
* gdc.dg/gdc36.d: Likewise.
* gdc.dg/gdc51.d: Likewise.
* gdc.dg/gdc57.d: Likewise.
* gdc.dg/gdc66.d: Likewise.
* gdc.dg/imports/gdc36.d: Likewise.
* gdc.dg/init1.d: Likewise.
* gdc.dg/pr92309.d: Likewise.
* gdc.dg/pr94424.d: Likewise.
* gdc.dg/pr94777b.d: Likewise.
* gdc.dg/pr96152.d: Likewise.
* gdc.dg/pr96153.d: Likewise.
* gdc.dg/pr96156.d: Likewise.
* gdc.dg/pr96157a.d: Likewise.
* gdc.dg/torture/torture.exp: New file.

c++: Stop defining true, false and bool as macros in <stdbool.h>

Since r216679 these macros have only been defined in C++98 mode, rather
than all modes. That is permitted as a GNU extension because that header
doesn't exist in the C++ standard until C++11, so we can make it do
whatever we want for C++98. But as discussed in the PR c++/60304
comments, these macros shouldn't ever be defined for C++.

This patch removes the macro definitions for C++98 too.

The new test already passed for C++98 (and the conversion is ill-formed
in C++11 and later) so this new test is arguably unnecessary.

gcc/ChangeLog:

PR c++/60304
* ginclude/stdbool.h (bool, false, true): Never define for C++.

gcc/testsuite/ChangeLog:

PR c++/60304
* g++.dg/warn/Wconversion-null-5.C: New test.

testsuite: Add missing <exception> header to testcase

This test no longer compiles because <new> stopped including
<exception>, so std::set_terminate is not defined.

gcc/testsuite/ChangeLog:

* g++.old-deja/g++.abi/cxa_vec.C: Include <exception> for
std::set_terminate.

libstdc++: Fix test to use correct function

This was copied from a test for std::lcm but I forgot to change one of
the calls to use the experimental version of the function.

libstdc++-v3/ChangeLog:

PR libstdc++/92978
* testsuite/experimental/numeric/92978.cc: Use experimental::lcm
not std::lcm.

MSP430: Fix -mlarge documentation to indicate size_t is a 20-bit type

gcc/ChangeLog:

* doc/invoke.texi (MSP430 options): Fix -mlarge description to
indicate size_t is a 20-bit type.

libstdc++: Fix three-way comparison for std::array [PR 96851]

The spaceship operator for std::array uses memcmp when the
__is_byte<value_type> trait is true, but memcmp isn't usable in
constexpr contexts. Also, memcmp should only be used for unsigned byte
types, because it gives the wrong answer for signed chars with negative
values.

We can simply check std::is_constant_evaluated() so that we don't use
memcmp during constant evaluation.

To fix the problem of using memcmp for inappropriate types, this patch
adds new __is_memcmp_ordered and __is_memcmp_ordered_with traits. These
say whether using memcmp will give the right answer for ordering
operations such as lexicographical_compare and three-way comparisons.
The new traits can be used in several places, and can also be used to
implement my suggestion in PR 93059 comment 37 to use memcmp for
unsigned integers larger than one byte on big endian targets.

libstdc++-v3/ChangeLog:

PR libstdc++/96851
* include/bits/cpp_type_traits.h (__is_memcmp_ordered):
New trait that says if memcmp can be used for ordering.
(__is_memcmp_ordered_with): Likewise, for two types.
* include/bits/deque.tcc (__lex_cmp_dit): Use new traits
instead of __is_byte and __numeric_traits.
(__lexicographical_compare_aux1): Likewise.
* include/bits/ranges_algo.h (__lexicographical_compare_fn):
Likewise.
* include/bits/stl_algobase.h (__lexicographical_compare_aux1)
(__is_byte_iter): Likewise.
* include/std/array (operator<=>): Likewise. Only use memcmp
when std::is_constant_evaluated() is false.
* testsuite/23_containers/array/comparison_operators/96851.cc:
New test.
* testsuite/23_containers/array/tuple_interface/get_neg.cc:
Adjust dg-error line numbers.

MSP430: Skip gcc.dg/pr55940.c in the small memory model

In the MSP430 small memory model, there is a 16-bit address space and
pointer arithmetic wraps around the address space, so any calculated
address is always within this range.

In this test, pointer arithmetic wraps when 0x1000 is added to the
address of a variable, causing the resulting address to be unexpectedly
less than 0x2000, which breaks the test.

gcc/testsuite/ChangeLog:

* gcc.dg/pr55940.c: Skip for msp430 unless -mlarge is specified.

libstdc++: Break header cycle between <new> and <exception>

The <new> and <exception> headers each include each other, which makes
building them as header-units "exciting". The <new> header only needs
the definition of std::exception (in order to derive from it) which is
already in its own header, so just include that.

libstdc++-v3/ChangeLog:

* include/bits/stl_iterator.h: Include <bits/exception_defines.h>
for definitions of __try, __catch and __throw_exception_again.
(counted_iterator::operator++(int)): Use __throw_exception_again
instead of throw.
* libsupc++/new: Include <bits/exception.h> not <exception>.
* libsupc++/new_opvnt.cc: Include <bits/exception_defines.h>.
* testsuite/18_support/destroying_delete.cc: Include
<type_traits> for std::is_same_v definition.
* testsuite/20_util/variant/index_type.cc: Qualify size_t.

fortran: Fix o'...' boz to integer/real conversions [PR96859]

The standard says that excess digits from boz are truncated.
For hexadecimal or binary, the routines copy just the number of digits
that will be needed, but for octal we copy number of digits that
contain one extra bit (for 8-bit, 32-bit or 128-bit, i.e. kind 1, 4 and 16)
or two extra bits (for 16-bit or 64-bit, i.e. kind 2 and 8).
The clearing of the first bit is done correctly by changing the first digit
if it is 4-7 to one smaller by 4 (i.e. modulo 4).
The clearing of the first two bits is done by changing 4 or 6 to 0
and 5 or 7 to 1, which is incorrect, because we really want to change the
first digit to 0 if it was even, or to 1 if it was odd, so digits
2 and 3 are mishandled by keeping them as is, rather than changing 2 to 0
and 3 to 1.

2020-09-02  Jakub Jelinek  <jakub@redhat.com>

PR fortran/96859
* check.c (gfc_boz2real, gfc_boz2int): When clearing first two bits,
change also '2' to '0' and '3' to '1' rather than just handling '4'
through '7'.

* gfortran.dg/pr96859.f90: New test.

hppa: Improve hppa_rtx_costs for shifts by constants.

This patch provides more accurate rtx_costs estimates for shifts by
integer constants (which are cheaper than by a register amount).

2020-09-02  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/pa/pa.c (hppa_rtx_costs) [ASHIFT, ASHIFTRT, LSHIFTRT]:
Provide accurate costs for shifts of integer constants.

bpf: use the default asm_named_section target hook

This patch makes the BPF backend to not provide its own implementation
of the asm_named_section hook; the default handler works perfectly
well.

2020-09-02  Jose E. Marchesi  <jose.marchesi@oracle.com>

gcc/
* config/bpf/bpf.c (bpf_asm_named_section): Delete.
(TARGET_ASM_NAMED_SECTION): Likewise.

bpf: use elfos.h

BPF is an ELF-based target, so it definitely benefits from using
elfos.h.  This patch makes the target to use it, and removes
superfluous definitions from bpf.h which are better defined in
elfos.h.

Note that BPF, despite being an ELF target, doesn't use DWARF.  At
some point it will generate DWARF when generating xBPF (-mxbpf) and
BTF when generating plain eBPF, but for the time being it just
generates stabs.

2020-09-02  Jose E. Marchesi  <jemarch@gnu.org>

gcc/
* config.gcc: Use elfos.h in bpf-*-* targets.
* config/bpf/bpf.h (MAX_OFILE_ALIGNMENT): Remove definition.
(COMMON_ASM_OP): Likewise.
(INIT_SECTION_ASM_OP): Likewise.
(FINI_SECTION_ASM_OP): Likewise.
(ASM_OUTPUT_SKIP): Likewise.
(ASM_OUTPUT_ALIGNED_COMMON): Likewise.
(ASM_OUTPUT_ALIGNED_LOCAL): Likewise.

Daily bump.

Use the determined lower bound of the range of offsets in a PLUS_EXPR.

gcc/ChangeLog:

* builtins.c (compute_objsize):  Only replace the upper bound
of a POINTER_PLUS offset when it's less than the lower bound.

gcc/testsuite/ChangeLog:

* gcc.dg/Wstringop-overflow.c: Remove xfails.
* gcc.dg/Wstringop-overflow-42.c: New test.
* gcc.dg/Wstringop-overread-4.c: New test.

c++: Allow new char[4]{"foo"} [PR77841]

Currently, we allow new char[]{"foo"}, but not new char[4]{"foo"}.
We should accept the latter too: [dcl.init.list]p3.3 says to treat
this as [dcl.init.string].

We were rejecting this code because we never called reshape_init before
the digest_init in build_new_1.  reshape_init handles [dcl.init.string]
by unwrapping the STRING_CST from its enclosing { }, and digest_init
assumes that reshape_init has been called for aggregates anyway, and an
array is an aggregate.

gcc/cp/ChangeLog:

PR c++/77841
* init.c (build_new_1): Call reshape_init.

gcc/testsuite/ChangeLog:

PR c++/77841
* g++.dg/cpp0x/initlist-new4.C: New test.

libstdc++: Add compile-time checks to__glibcxx_assert [PR 71960]

This change evaluates __glibcxx_assert checks unconditionally when a
function is being constant evaluated (when std::is_constant_evaluated()
is true). If the check fails, compilation will fail with an error.

If the function isn't being constant evaluated, the normal runtime check
will be done if enabled by _GLIBCXX_ASSERTIONS or _GLIBCXX_DEBUG, the
same as before.

Tangentially, the __glibcxx_assert and _GLIBCXX_PARALLEL_ASSERT macros
are changed to expand to 'do { } while (false)' when assertions are
disabled, instead of expanding to nothing. This avoids -Wempty-body
warnings when a disabled assertion is used in an 'if' or 'else'
statement e.g.

  if constexpr (/* precondition is testable */)
    __glibcxx_assert(precondition);

a.C:9:27: warning: suggest braces around empty body in an ‘if’ statement [-Wempty-body]
    9 |     __glibcxx_assert(precondition);
      |                                  ^

libstdc++-v3/ChangeLog:

PR libstdc++/71960
* include/bits/c++config (__glibcxx_assert_impl): Remove
do-while so that uses of the macro need to add it.
(__glibcxx_assert): Rename macro for runtime assertions
to __glibcxx_assert_2.
(__glibcxx_assert_1): Define macro for constexpr assertions.
(__glibcxx_assert): Define macro for constexpr and runtime
assertions.
* include/bits/range_access.h (ranges::advance): Remove
redundant precondition checks during constant evaluation.
* include/parallel/base.h (_GLIBCXX_PARALLEL_ASSERT): Always
use do-while in macro expansion.
* include/std/ranges (iota_view::iota_view(W, B)): Remove
redundant braces.

rs6000: MMA built-in dies with incorrect sharing of tree nodes error

When we expand our MMA built-ins into gimple, we erroneously reused the
accumulator memory reference for both the source input value as well as
the destination output value.  This led to a tree sharing error.
The solution is to create separate memory references for the input
and output values.

2020-09-01  Peter Bergner  <bergner@linux.ibm.com>

gcc/
PR target/96808
* config/rs6000/rs6000-call.c (rs6000_gimple_fold_mma_builtin): Do not
reuse accumulator memory reference for source and destination accesses.

gcc/testsuite/
PR target/96808
* gcc.target/powerpc/pr96808.c: New test.

libstdc++: Constrain chrono::duration conversions [LWG 2094]

The chrono::duration constructor that converts from another duration
type is meant to be constrained so that it doesn't participate in
overload resolution if the ratio of the periods cannot be represented as
a std::ratio.

Because our std::ratio_divide is not SFINAE-friendly the evaluation of
__is_harmonic results in an error outside the immediate context when an
overflow occurs. I intend to make ratio_divide (and ratio_multiply)
SFINAE-friendly in a future patch, but for now this patch just
introduces a new SFINAE-friendly alias template for the division.

The standard doesn't require it, but it also seems right to constrain
the constructor with std::is_convertible_v<_Rep2, rep>.

libstdc++-v3/ChangeLog:

* include/std/chrono (duration::_S_gcd(intmax_t, intmax_t)):
New helper function for finding GCD of two positive intmax_t
values.
(duration::__divide): New helper alias for dividing one period
by another.
(duration::__is_harmonic): Use __divide not ratio_divide.
(duration(const duration<R2, P2>&)): Require the duration rep
types to be convertible.
* testsuite/20_util/duration/cons/dr2094.cc: New test.
* testsuite/20_util/duration/requirements/reduced_period.cc:
Fix definition of unused member functions in test type.
* testsuite/20_util/duration/requirements/typedefs_neg2.cc:
Adjust expected errors.

analyzer: fix false NULL deref warning after previous deref [PR96792]

gcc/analyzer/ChangeLog:
PR analyzer/96792
* region-model.cc (region_model::deref_rvalue): Add the constraint
that PTR_SVAL is non-NULL.

gcc/testsuite/ChangeLog:
PR analyzer/96792
* gcc.dg/analyzer/pr96792.c: New test.

vec: use inexact growth where possible.

gcc/ChangeLog:

* cfgrtl.c (rtl_create_basic_block): Use default value for
growth vector function.
* gimple.c (gimple_set_bb): Likewise.
* symbol-summary.h: Likewise.
* tree-cfg.c (init_empty_tree_cfg_for_function): Likewise.
(build_gimple_cfg): Likewise.
(create_bb): Likewise.
(move_block_to_fn): Likewise.

vec: default exect = false in grow functions.

gcc/ChangeLog:

* vec.h (vec_safe_grow): Change default of exact to false.
(vec_safe_grow_cleared): Likewise.

PR middle-end/90597: gcc_assert ICE in layout_type

This patch fixes the default implementation of TARGET_VECTOR_ALIGNMENT,
known as default_vector_alignment, using the same logic as my earlier
nvptx patch, as the ICE caused by TYPE_SIZE(type) being zero during
error handling in gcc.dg/attr-vector_size.c is common among backends,
and is known in bugzilla as PR middle-end/90597, apparently a recent
regression.

2020-09-01  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog:
PR middle-end/90597
* targhooks.c (default_vector_alignment): Return at least the
GET_MODE_ALIGNMENT for the type's mode.

Fortran  : ICE on invalid code PR95398

The CLASS_DATA macro is used to shorten the code accessing the derived
components of an expressions type specification.  If the type is not
BT_CLASS the derived pointer is NULL resulting in an ICE.  To avoid
dereferencing a NULL pointer the type should be BT_CLASS.

2020-09-01  Steven G. Kargl  <kargl@gcc.gnu.org>

gcc/fortran

PR fortran/95398
* resolve.c (resolve_select_type): Add check for BT_CLASS
type before using the CLASS_DATA macro which will have a
NULL pointer to derive components if it isn't BT_CLASS.

2020-09-01  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite

PR fortran/95398
* gfortran.dg/pr95398.f90: New test.

rtl-optimization/96812 - remap dependence info on RTL loop unrolling

This carries over the PR87609 fix also to RTL loop unrolling.  The
gcc.dg/torture/pr90328.c testcase otherwise is miscompiled with
the tree-ssa-address.c hunk (or alternatively with -fno-ivopts
on master).  I've tried to find the correct abstraction and
adjusted two other duplicate_insn_chain users for which I do not
have testcases.  There may be other insn-chain copying routines
that could be affected but hopefully most appropriately go through
CFG hooks.

2020-08-27  Richard Biener  <rguenther@suse.de>

PR rtl-optimization/96812
* tree-ssa-address.c (copy_ref_info): Also copy dependence info.
* cfgrtl.h (duplicate_insn_chain): Adjust prototype.
* cfgrtl.c (duplicate_insn_chain): Remap dependence info
if requested.
(cfg_layout_duplicate_bb): Make sure we remap dependence info.
* modulo-sched.c (duplicate_insns_of_cycles): Remap dependence
info.
(generate_prolog_epilog): Adjust.
* config/c6x/c6x.c (hwloop_optimize): Remap dependence info.

openmp: Check for PARM_DECL before using C_ARRAY_PARAMETER or DECL_ARRAY_PARAMETER_P [PR96867]

The C++ macro performs a PARM_DECL_CHECK, so will ICE if not tested on a PARM_DECL,
C_ARRAY_PARAMETER doesn't, but probably should, otherwise it is testing e.g.
C_DECL_VARIABLE_SIZE on VAR_DECLs.

2020-09-01  Jakub Jelinek  <jakub@redhat.com>

PR c++/96867
* c-typeck.c (handle_omp_array_sections_1): Test C_ARRAY_PARAMETER
only on PARM_DECLs.

* semantics.c (handle_omp_array_sections_1): Test
DECL_ARRAY_PARAMETER_P only on PARM_DECLs.

* c-c++-common/gomp/pr96867.c: New test.

Correct Changelog for fix to PR96806

test/rs6000: Add Power9 and up as vect_len target

Power9 supports vector with length in bytes load/store, this patch
is to teach check_effective_target_vect_len_load_store to take it
and its laters as effective vector with length targets.

Also supplement the documents for has_arch_pwr*.

Bootstrapped/regtested on powerpc64le-linux-gnu P8, also on
powerpc64le-linux-gnu P9 with explicit usage setting.

gcc/ChangeLog:

* doc/sourcebuild.texi (has_arch_pwr5, has_arch_pwr6, has_arch_pwr7,
has_arch_pwr8, has_arch_pwr9): Document.

gcc/testsuite/ChangeLog:

* lib/target-supports.exp
(check_effective_target_vect_len_load_store): Call check function
check_effective_target_has_arch_pwr9.

Daily bump.

rs6000, remove improperly defined and unsupported builtins.

gcc/ChangeLog

2020-08-31  Carl Love  <cel@us.ibm.com>

PR target/85830
* config/rs6000/altivec.h (vec_popcntb, vec_popcnth, vec_popcntw,
vec_popcntd): Remove defines.

analyzer: handle __builtin___memset_chk [PR96798]

gcc/analyzer/ChangeLog:
PR analyzer/96798
* region-model.cc (region_model::on_call_pre): Handle
BUILT_IN_MEMSET_CHK.

gcc/testsuite/ChangeLog:
PR analyzer/96798
* gcc.dg/analyzer/memset-1.c (test_5a): New.

analyzer: gather builtin/internal fn handling into switch statements

Clean up this code in preparation for fixing PR analyzer/96798.

gcc/analyzer/ChangeLog:
* region-model.cc (region_model::on_call_pre): Gather handling of
builtins and of internal fns into switch statements.  Handle
"alloca" and BUILT_IN_ALLOCA_WITH_ALIGN.

analyzer: fix ICE on unknown index in CONSTRUCTOR [PR96860]

PR analyzer/96860 reports an ICE inside CONSTRUCTOR-handling with
--param analyzer-max-svalue-depth=0 when attempting to build a
binding_map for the CONSTRUCTOR's values.

The issue is that when handling (index, value) pairs for initializing
an array, the index values for the elements exceeds the svalue
complexity limit, and the index is thus treated as unknown, leading to
a symbolic rather than concrete offset for each array element.

This patch updates the CONSTRUCTOR-handling code so that it can
fail, returning an unknown value for the overall value of the
constructor for this case, fixing the ICE.

gcc/analyzer/ChangeLog:
PR analyzer/96860
* region.cc (decl_region::get_svalue_for_constructor): Support
apply_ctor_to_region failing.
* store.cc (binding_map::apply_ctor_to_region): Add failure
handling.
(binding_map::apply_ctor_val_to_range): Likewise.
(binding_map::apply_ctor_pair_to_child_region): Likewise.  Replace
assertion that child_base_offset is not symbolic with error
handling.
* store.h (binding_map::apply_ctor_to_region): Convert return type
from void to bool.
(binding_map::apply_ctor_val_to_range): Likewise.
(binding_map::apply_ctor_pair_to_child_region): Likewise.

gcc/testsuite/ChangeLog:
PR analyzer/96860
* gcc.dg/analyzer/pr96860-1.c: New test.
* gcc.dg/analyzer/pr96860-2.c: New test.

c++: Implement P1009: Array size deduction in new-expressions.

This patch implements C++20 P1009, allowing code like

  new double[]{1,2,3}; // array bound will be deduced

Since this proposal makes the initialization rules more consistent, it is
applied to all previous versions of C++ (thus, effectively, all the way back
to C++11).

My patch is based on Jason's patch that handled the basic case.  I've
extended it to work with ()-init and also the string literal case.
Further testing revealed that to handle stuff like

  new int[]{t...};

in a template, we have to consider such a NEW_EXPR type-dependent.
Obviously, we first have to expand the pack to be able to deduce the
number of elements in the array.

Curiously, while implementing this proposal, I noticed that we fail
to accept

  new char[4]{"abc"};

so I've assigned 77841 to self.  I think the fix will depend on the
build_new_1 hunk in this patch.

The new tree.c function build_constructor_from_vec helps us morph
a vector into a CONSTRUCTOR more efficiently.

gcc/cp/ChangeLog:

PR c++/93529
* call.c (build_new_method_call_1): Use build_constructor_from_vec
instead of build_tree_list_vec + build_constructor_from_list.
* init.c (build_new_1): Handle new char[]{"foo"}.  Use
build_constructor_from_vec instead of build_tree_list_vec +
build_constructor_from_list.
(build_new): Deduce the array size in new-expression if not
present.  Handle ()-init.  Handle initializing an array from
a string literal.
* parser.c (cp_parser_new_type_id): Leave [] alone.
(cp_parser_direct_new_declarator): Allow [].
* pt.c (type_dependent_expression_p): In a NEW_EXPR, consider
array types whose dimension has to be deduced type-dependent.

gcc/ChangeLog:

PR c++/93529
* tree.c (build_constructor_from_vec): New.
* tree.h (build_constructor_from_vec): Declare.

gcc/testsuite/ChangeLog:

PR c++/93529
* g++.dg/cpp0x/sfinae4.C: Adjust expected result after P1009.
* g++.dg/cpp2a/new-array1.C: New test.
* g++.dg/cpp2a/new-array2.C: New test.
* g++.dg/cpp2a/new-array3.C: New test.
* g++.dg/cpp2a/new-array4.C: New test.

Co-authored-by: Jason Merrill <jason@redhat.com>

analyzer: fix ICE on RANGE_EXPR in CONSTRUCTORs [PR96763]

gcc/analyzer/ChangeLog:
PR analyzer/96763
* store.cc (binding_map::apply_ctor_to_region): Handle RANGE_EXPR
by calling a new binding_map::apply_ctor_val_to_range subroutine.
Split out the existing non-CONSTRUCTOR-handling code to a new
apply_ctor_pair_to_child_region subroutine.
(binding_map::apply_ctor_val_to_range): New.
(binding_map::apply_ctor_pair_to_child_region): New, split out
from binding_map::apply_ctor_to_region as noted above.
* store.h (binding_map::apply_ctor_val_to_range): New decl.
(binding_map::apply_ctor_pair_to_child_region): New decl.

gcc/testsuite/ChangeLog:
PR analyzer/96763
* g++.dg/analyzer/pr96763.C: New test.

analyzer: fix ICE on casting float to pointer [PR96764]

gcc/analyzer/ChangeLog:
PR analyzer/96764
* region-model-manager.cc
(region_model_manager::maybe_fold_unaryop): Handle VIEW_CONVERT_EXPR.
(region_model_manager::get_or_create_cast): Move logic for
real->integer casting to...
(get_code_for_cast): ...this new function, and add logic for
real->non-integer casts.
(region_model_manager::maybe_fold_sub_svalue): Handle
VIEW_CONVERT_EXPR.
* region-model.cc
(region_model::add_any_constraints_from_gassign): Likewise.
* svalue.cc (svalue::maybe_undo_cast): Likewise.
(unaryop_svalue::dump_to_pp): Likewise.

gcc/testsuite/ChangeLog:
PR analyzer/96764
* gcc.dg/analyzer/pr96764.c: New test.

PR tree-optimization/96818 - cast label range to type of switch operand

PR tree-optimization/96818
* tree-vrp.c (find_case_label_range): Cast label range to
type of switch operand.

d: Fix ICEs in the front-end when pointer size is 16-bit.

In the lowering of `bt*' intrinsics, some integer constants had
mismatched types, and bitsize was set to the wrong value.

In base_vtable_offset, the base offset value was calculated incorrectly.
The TypeInfo_Class object is comprised of 18 pointers and 1 uint field,
so now the internal classinfo type size is used instead.

gcc/d/ChangeLog:

* d-target.cc (Target::_init): Don't set classinfosize.
* d-tree.h (base_vtable_offset): Move under typeinfo.cc section.
* decl.cc (base_vtable_offset): Move to...
* typeinfo.cc (base_vtable_offset): ...here.  Get base offset from
internal TypeInfo_Class type.
* intrinsics.cc (expand_intrinsic_bt): Use pointer TYPE_SIZE for
setting bitsize value.  Build integer constants of correct type.

tree-optimization/96854 - testcase for SLP reduction of two-operator

This adds the testcase for the already fixed PR.

2020-08-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96854
* gcc.dg/vect/pr96854.c: New testcase.

testsuite: Fix aarch64/strcmpopt_6.c for -Wstringop-overread

2020-08-31  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/testsuite/
* gcc.target/aarch64/strcmpopt_6.c: Suppress -Wstringop-overread.

Refine expander vec_unpacku_float_hi_v16si/vec_unpacku_float_lo_v16si

gcc/
PR target/96551
* config/i386/sse.md (vec_unpacku_float_hi_v16si): For vector
compare to integer mask, don't use gen_rtx_LT, use
ix86_expand_mask_vec_cmp instead.
(vec_unpacku_float_hi_v16si): Ditto.

gcc/testsuite
* gcc.target/i386/avx512f-pr96551-1.c: New test.
* gcc.target/i386/avx512f-pr96551-2.c: New test.

tree-cfg: Improve gimple switch verification

When looking at the verification, I have noticed a bug in it.

The verification that CASE_HIGH (if present) has the same type as CASE_LOW
is only performed for the case label 2 and higher, case label 1 (the first
one after the default label) isn't checked.

The following patch fixes that, it will uselessly also compare
TREE_TYPE (CASE_LOW (elt)) != elt_type for the case label 1, but I think
that isn't that expensive and helps readability of the code.

2020-08-31  Jakub Jelinek  <jakub@redhat.com>

* tree-cfg.c (verify_gimple_switch): If the first non-default case
label has CASE_HIGH, verify it has the same type as CASE_LOW.

ipa/96806 - Fix ICE in ipa-cp due to integer addition overflow

2020-08-31  Feng Xue  <fxue@os.amperecomputing.com>

gcc/
PR tree-optimization/96806
* ipa-cp.c (decide_about_value): Use safe_add to avoid cost addition
overflow.

gcc/testsuite/
PR tree-optimization/96806
* g++.dg/ipa/pr96806.C: New test.

varasm: Optimize memory broadcast for constant vector under AVX512 [PR54201]

I meant something like the following, which on e.g. a dumb:

typedef float V __attribute__((vector_size (4 * sizeof (float))));

void
foo (V *p, float *q)
{
  p[0] += (V) { 1.0f, 2.0f, 3.0f, 4.0f };
  q[0] += 4.0f;
  q[1] -= 3.0f;
  q[17] -= 2.0f;
  q[31] += 1.0f;
}

testcase merges all the 4 scalar constant pool entries into the CONST_VECTOR
one.

I'm punting for section anchors and not doing it in the per-function (i.e.
non-shared) constant pools simply because I don't know them well enough,
don't know whether backends use the offsets for something etc.
For section anchors, I guess it would need to be done before (re)computing the
offsets and arrange for the desc->mark < 0 entries not to be considered as
objects in the object block, for non-shared pools, perhaps it would be
enough to call the new function from output_constant_pool before calling
recompute_pool_offsets and adjust recompute_pool_offsets to ignore
desc->mark < 0.

Here is an adjusted patch that ought to merge even the same sized different
mode vectors with the same byte representation, etc.
It won't really help with avoiding the multiple reads of the constant in the
same function, but as you found, your patch doesn't help with that either.
Your patch isn't really incompatible with what the patch below does, though
I wonder whether a) it wouldn't be better to always canonicalize to an
integral mode with as few elts as possible even e.g. for floats b) whether
asserting that it simplify_rtx succeeds is safe, whether it shouldn't just
canonicalize if the canonicalization works and just do what it previously
did otherwise.

The following patch puts all pool entries which can be natively encoded
into a vector, sorts it by decreasing size, determines minimum size
of a pool entry and adds hash elts for each (aligned) min_size or wider
power of two-ish portion of the pool constant in addition to the whole pool
constant byte representation.

This is the version that passed bootstrap/regtest on both x86_64-linux and
i686-linux.  In both bootstraps/regtests together, it saved (from the
statistics I've gathered) 63104 .rodata bytes (before constant merging),
in 6814 hits of the data->desc->mark = ~(*slot)->desc->labelno;.

2020-08-31  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/54201
* varasm.c: Include alloc-pool.h.
(output_constant_pool_contents): Emit desc->mark < 0 entries as
aliases.
(struct constant_descriptor_rtx_data): New type.
(constant_descriptor_rtx_data_cmp): New function.
(struct const_rtx_data_hasher): New type.
(const_rtx_data_hasher::hash, const_rtx_data_hasher::equal): New
methods.
(optimize_constant_pool): New function.
(output_shared_constant_pool): Call it if TARGET_SUPPORTS_ALIASES.

testsuite: Update some vect cases for partial vectors

This patch is to adjust some existing vectorization test cases
to work well with the newly introduced partial vector usages.

Bootstrapped/regtested on aarch64-linux-gnu and powerpc64le-linux-gnu
P9 (with explicit param vect-partial-vector-usage=1 and enablement on
check_effective_target_vect_partial_vectors_usage_1 check).

gcc/ChangeLog:

* doc/sourcebuild.texi (vect_len_load_store,
vect_partial_vectors_usage_1, vect_partial_vectors_usage_2,
vect_partial_vectors): Document.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/bb-slp-pr69907.c: Adjust for partial vector usages.
* gcc.dg/vect/slp-3.c: Likewise.
* gcc.dg/vect/slp-multitypes-11.c: Likewise.
* gcc.dg/vect/slp-perm-1.c: Likewise.
* gcc.dg/vect/slp-perm-5.c: Likewise.
* gcc.dg/vect/slp-perm-6.c: Likewise.
* gcc.dg/vect/slp-perm-7.c: Likewise.
* gcc.dg/vect/slp-perm-8.c: Likewise.
* gcc.dg/vect/slp-perm-9.c: Likewise.
* gcc.dg/vect/vect-version-2.c: Likewise.
* lib/target-supports.exp (check_vect_partial_vector_usage): New
function.
(check_effective_target_vect_len_load_store): Likewise.
(check_effective_target_vect_partial_vectors_usage_1): Likewise.
(check_effective_target_vect_partial_vectors_usage_2): Likewise.
(check_effective_target_vect_partial_vectors): Likewise.

Daily bump.

Use get_size_range instead of get_range to obtain range of valid sizes.

gcc/ChangeLog:

* builtins.c (access_ref::access_ref): Call get_size_range instead
of get_range.

gcc/testsuite/ChangeLog:

* gcc.dg/Wstringop-overread-3.c: New test.

2020-8-20  José Rui Faustino de Sousa  <jrfsousa@gmail.com>

gcc/fortran/ChangeLog:

PR fortran/96728
* module.c (module_peek_char): Peek ahead function.
(parse_integer): Add code for parsing signed integers.
(parse_atom): Add code to handle signed integers.
(peek_atom): Add code to handle signed integers.

gcc/testsuite/ChangeLog:

PR fortran/96728
* gfortran.dg/PR96728.f90: New test.

2020-8-20  José Rui Faustino de Sousa  <jrfsousa@gmail.com>

gcc/fortran/ChangeLog:

PR fortran/96727
* expr.c (gfc_check_init_expr): Add default error message for the AS_ASSUMED_RANK case.

gcc/testsuite/ChangeLog:

PR fortran/96727
* gfortran.dg/PR96727.f90: New test.

2020-8-20  José Rui Faustino de Sousa  <jrfsousa@gmail.com>

gcc/fortran/ChangeLog:

PR fortran/96726
* expr.c (check_references): Change different relational operator
to less-than operator to avoid infinite loop.

gcc/testsuite/ChangeLog:

PR fortran/96726
* gfortran.dg/PR96726.f90: New test.

2020-8-21  Steve Kargl <sgk@troutmask.apl.washington.edu>

gcc/fortran/ChangeLog:

PR fortran/95352
* simplify.c (simplify_bound_dim): Add check for NULL pointer
before trying to access structure member.

    José Rui Faustino de Sousa  <jrfsousa@gmail.com>

gcc/testsuite/ChangeLog:

* gfortran.dg/PR95352.f90: New test.

2020-8-20  José Rui Faustino de Sousa  <jrfsousa@gmail.com>

gcc/fortran/ChangeLog:

PR fortran/94110
* interface.c (gfc_compare_actual_formal): Add code to also raise
the actual argument cannot be an assumed-size array error when the
dummy arguments are deferred-shape or assumed-rank pointer.

gcc/testsuite/ChangeLog:

PR fortran/94110
* gfortran.dg/PR94110.f90: New test.

x86: Fix up ssse3_pshufbv8qi splitter

The constant pool size optimization I was testing resulted in various ICEs
in gcc.target/i386/ testsuite, the problem is that the ssse3_pshufbv8qi
splitter emits invalid RTL, in V4SImode 0xf7f7f7f7 CONST_INTs shouldn't
appear, instead they should have been -0x8080809 (0xf7f7f7f7 sign extended
into 64 bits).

2020-08-30  Jakub Jelinek  <jakub@redhat.com>

* config/i386/sse.md (ssse3_pshufbv8qi): Use gen_int_mode instead of
GEN_INT, and ix86_build_const_vector instead of gen_rtvec and
gen_rtx_CONT_VECTOR.

Daily bump.

libstdc++: Fix deleted overload of __absu(bool)

libstdc++-v3/ChangeLog:

* include/std/numeric (__detail::__absu(bool)): Make deleted
function a function template, so it will be chosen for calls
with an explicit template argument list.
* testsuite/26_numerics/gcd/gcd_neg.cc: Add dg-prune-output.
* testsuite/26_numerics/lcm/lcm_neg.cc: Likewise.

rs6000: Remove ALTIVEC_BUILTIN_MASK_FOR_STORE

It turns out that the target hook that this is supposed to satisfy
disappeared in 2004.  Probably time to retire it.

2020-08-28  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin.def (MASK_FOR_STORE): Remove.
* config/rs6000/rs6000-call.c (rs6000_expand_builtin): Remove
all logic for ALTIVEC_BUILTIN_MASK_FOR_STORE.

Daily bump.

libstdc++: Fix common_type specializations for duration

My recent change to implement P0548 ("common_type and duration") was not
correct. The result of common_type_t<duration<R,P>, duration<R,P>>
should be duration<common_type_t<R>, P::type>, not duration<R, P::type>.
The common_type specialization for two different duration types was
correct, but the specializations for a single duration type (which only
exist to optimize compilation time) were wrong.

This fixes the partial specializations of common_type for a single
duration type, and also the return types of duration::operator+ and
duration::operator- which are supposed to use common_type_t<duration>.

libstdc++-v3/ChangeLog:

* include/std/chrono (common_type): Fix partial specializations
for a single duration type to use the common_type of the rep.
(duration::operator+, duration::operator-): Fix return types
to also use the common_type of the rep.
* testsuite/20_util/duration/requirements/reduced_period.cc:
Check duration using a rep that has common_type specialized.

libstdc++: Fix std::gcd and std::lcm for unsigned integers [PR 92978]

This fixes a bug with mixed signed and unsigned types, where converting
a negative value to the unsigned result type alters the value. The
solution is to obtain the absolute values of the arguments immediately
and to perform the actual GCD or LCM algorithm on two arguments of the
same type.

In order to operate on the most negative number without overflow when
taking its absolute, use an unsigned type for the result of the abs
operation. For example, -INT_MIN will overflow, but -(unsigned)INT_MIN
is (unsigned)INT_MAX+1U which is the correct value.

libstdc++-v3/ChangeLog:

PR libstdc++/92978
* include/std/numeric (__abs_integral): Replace with ...
(__detail::__absu): New function template that returns an
unsigned type, guaranteeing it can represent the most
negative signed value.
(__detail::__gcd, __detail::__lcm): Require arguments to
be unsigned and therefore already non-negative.
(gcd, lcm): Convert arguments to absolute value as unsigned
type before calling __detail::__gcd or __detail::__lcm.
* include/experimental/numeric (gcd, lcm): Likewise.
* testsuite/26_numerics/gcd/gcd_neg.cc: Adjust expected
errors.
* testsuite/26_numerics/lcm/lcm_neg.cc: Likewise.
* testsuite/26_numerics/gcd/92978.cc: New test.
* testsuite/26_numerics/lcm/92978.cc: New test.
* testsuite/experimental/numeric/92978.cc: New test.

Add -Wstringop-overread for reading past the end by string functions.

gcc/ChangeLog:

* attribs.c (init_attr_rdwr_indices): Use global access_mode.
* attribs.h (struct attr_access): Same.
* builtins.c (fold_builtin_strlen): Add argument.
(compute_objsize): Declare.
(get_range): Declare.
(check_read_access): New function.
(access_ref::access_ref): Define ctor.
(warn_string_no_nul): Add arguments.  Handle -Wstrintop-overread.
(check_nul_terminated_array): Handle source strings of different
ranges of sizes.
(expand_builtin_strlen): Remove warning code, call check_read_access
instead.  Declare locals closer to their initialization.
(expand_builtin_strnlen): Same.
(maybe_warn_for_bound): New function.
(warn_for_access): Remove argument.  Handle -Wstrintop-overread.
(inform_access): Change argument type.
(get_size_range): New function.
(check_access): Remove unused arguments.  Add new arguments.  Handle
-Wstrintop-overread.  Move warning code to helpers and call them.
Call check_nul_terminated_array.
(check_memop_access): Remove unnecessary and provide additional
arguments in calls.
(expand_builtin_memchr): Call check_read_access.
(expand_builtin_strcat): Remove unnecessary and provide additional
arguments in calls.
(expand_builtin_strcpy): Same.
(expand_builtin_strcpy_args): Same.  Avoid testing no-warning bit.
(expand_builtin_stpcpy_1): Remove unnecessary and provide additional
arguments in calls.
(expand_builtin_stpncpy): Same.
(check_strncat_sizes): Same.
(expand_builtin_strncat): Remove unnecessary and provide additional
arguments in calls.  Adjust comments.
(expand_builtin_strncpy): Remove unnecessary and provide additional
arguments in calls.
(expand_builtin_memcmp): Remove warning code.  Call check_access.
(expand_builtin_strcmp): Call check_access instead of
check_nul_terminated_array.
(expand_builtin_strncmp): Handle -Wstrintop-overread.
(expand_builtin_fork_or_exec): Call check_access instead of
check_nul_terminated_array.
(expand_builtin): Same.
(fold_builtin_1): Pass additional argument.
(fold_builtin_n): Same.
(fold_builtin_strpbrk): Remove calls to check_nul_terminated_array.
(expand_builtin_memory_chk): Add comments.
(maybe_emit_chk_warning): Remove unnecessary and provide additional
arguments in calls.
(maybe_emit_sprintf_chk_warning): Same.  Adjust comments.
* builtins.h (warn_string_no_nul): Add arguments.
(struct access_ref): Add member and ctor argument.
(struct access_data): Add members and ctor.
(check_access): Adjust signature.
* calls.c (maybe_warn_nonstring_arg): Return an indication of
whether a warning was issued.  Issue -Wstrintop-overread instead
of -Wstringop-overflow.
(append_attrname): Adjust to naming changes.
(maybe_warn_rdwr_sizes): Same.  Remove unnecessary and provide
additional arguments in calls.
* calls.h (maybe_warn_nonstring_arg): Return bool.
* doc/invoke.texi (-Wstringop-overread): Document new option.
* gimple-fold.c (gimple_fold_builtin_strcpy): Provide an additional
argument in call.
(gimple_fold_builtin_stpcpy): Same.
* tree-ssa-uninit.c (maybe_warn_pass_by_reference): Adjust to naming
changes.
* tree.h (enum access_mode): New type.

gcc/c-family/ChangeLog:

* c.opt (Wstringop-overread): New option.

gcc/testsuite/ChangeLog:

* c-c++-common/Warray-bounds-7.c: Adjust expected warnings.
* c-c++-common/Wrestrict.c: Remove xfail.
* c-c++-common/attr-nonstring-3.c: Adjust text of expected warnings.
* c-c++-common/attr-nonstring-6.c: Suppress -Wstringop-overread
instead of -Wstringop-overflow.
* c-c++-common/attr-nonstring-8.c: Adjust text of expected warnings.
* g++.dg/torture/Wsizeof-pointer-memaccess1.C: Also suppress
 -Wstringop-overread.
* g++.dg/torture/Wsizeof-pointer-memaccess2.C: Same.
* gcc.dg/Warray-bounds-39.c: Adjust expected warnings.
* gcc.dg/Warray-bounds-40.c: Also suppress -Wstringop-overread.
* gcc.dg/Warray-bounds-58.c: Remove xfail.  Also expect
-Wstringop-overread.  Adjust text of expected warnings.
* gcc.dg/Wsizeof-pointer-memaccess1.c: Also suppress
 -Wstringop-overread.
* gcc.dg/Wstringop-overflow-22.c: Adjust text of expected warnings.
* gcc.dg/Wstringop-overflow-33.c: Expect -Wstringop-overread.
* gcc.dg/Wstringop-overflow-9.c: Expect -Wstringop-overread.
* gcc.dg/attr-nonstring-2.c: Adjust text of expected warnings.
* gcc.dg/attr-nonstring-3.c: Same.
* gcc.dg/attr-nonstring-4.c: Same.
* gcc.dg/attr-nonstring.c: Expect -Wstringop-overread.
* gcc.dg/builtin-stringop-chk-5.c: Adjust comment.
* gcc.dg/builtin-stringop-chk-8.c: Enable -Wstringop-overread instead
of -Wstringop-overflow.
* gcc.dg/pr78902.c: Also expect -Wstringop-overread.
* gcc.dg/pr79214.c: Adjust text of expected warnings.
* gcc.dg/strcmpopt_10.c: Suppress valid -Wno-stringop-overread.
* gcc.dg/strlenopt-57.c: Also expect -Wstringop-overread.
* gcc.dg/torture/Wsizeof-pointer-memaccess1.c: Also suppress valid
-Wno-stringop-overread.
* gcc.dg/tree-ssa/builtins-folding-gimple-ub.c: Same.
* gcc.dg/uninit-33.c: Same.
* gcc.dg/warn-strnlen-no-nul-2.c: Adjust text of expected warning.
* gcc.dg/warn-strnlen-no-nul.c: Same.
* gcc.target/i386/strcmpopt_6.c: Suppress -Wstringop-overread.
* gcc.dg/Wstringop-overread-2.c: New test.
* gcc.dg/Wstringop-overread.c: New test.

rs6000: r12 copy cleanup

Remove unnecessary tests before copying function address to r12.

2020-08-28  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000.c (rs6000_call_aix): Remove test for r12.
(rs6000_sibcall_aix): Likewise.

amdgcn: Update vec_safe_grow_cleared usage

An API change broke the amdgcn build.

gcc/ChangeLog:

* config/gcn/gcn-tree.c (gcn_goacc_get_worker_red_decl): Add "true"
parameter to vec_safe_grow_cleared.

Correct calls to vec::safe_grow in conditionally compiled code.

gcc/ChangeLog:

* ggc-common.c (gt_pch_save): Add argument to a call.

gcc/jit/ChangeLog:

* jit-recording.c (recording::switch_::make_debug_string): Add argument
to a call.

PR c/96596 - ICE in match_builtin_function_types on a declaration of a built-in with invalid array argument

gcc/c/ChangeLog:

PR c/96596
* c-decl.c (match_builtin_function_types): Avoid dealing with erroneous
argument type.

gcc/testsuite/ChangeLog:

PR c/96596
* gcc.dg/Wbuiltin-declaration-mismatch-16.c: New test.

Fortran: Fix absent-optional handling for nondescriptor arrays (PR94672)

gcc/fortran/ChangeLog:

PR fortran/94672
* trans-array.c (gfc_trans_g77_array): Check against the parm decl and
set the nonparm decl used for the is-present check to NULL if absent.

gcc/testsuite/ChangeLog:

PR fortran/94672
* gfortran.dg/optional_assumed_charlen_2.f90: New test.

[PATCH PR96357][GCC][AArch64]: could not split insn UNSPEC_COND_FSUB with AArch64 SVE

Problem is related to that operand 4 (In original pattern
cond_sub<mode>_any_const) is no longer the same as operand 1, and so
the pattern doesn't match the split condition.

Pattern cond_sub<mode>_any_const is being split by this patch into two
separate patterns:
* Pattern cond_sub<mode>_relaxed_const now matches const_int
  SVE_RELAXED_GP operand.
* Pattern cond_sub<mode>_strict_const now matches const_int
  SVE_STRICT_GP operand.
* Remove aarch64_sve_pred_dominates_p condition from both patterns.

gcc/ChangeLog:

PR target/96357
* config/aarch64/aarch64-sve.md
(cond_sub<mode>_relaxed_const): Updated and renamed from
cond_sub<mode>_any_const pattern.
(cond_sub<mode>_strict_const): New pattern.

gcc/testsuite/ChangeLog:

PR target/96357
* gcc.target/aarch64/sve/pr96357.c: New test.

fix a typo in rtl.texi

gcc/
* doc/rtl.texi: Fix typo.

aarch64: Disable mem-shift-canonical test on ILP32

This test fails on ILP32 since we're looking for a pattern that could
only be hit on LP64. Disabling the test on ILP32 since the problematic
mult pattern was never hit there, so there's nothing to test.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/mem-shift-canonical.c: Skip on ILP32.

This patch fixes PR96624.

2020-08-28  Paul Thomas  <pault@gcc.gnu.org>

gcc/fortran
PR fortran/96624
* simplify.c (gfc_simplify_reshape): Detect zero shape and
clear index if found.

gcc/testsuite/
PR fortran/96624
* gfortran.dg/reshape_8.f90 : New test.

testsuite: Run gcc.dg/pr96579.c only on targets with dfp support.

gcc.dg/pr96579.c includes gcc.dg/pr96370.c which needs target dfp.

2020-08-28  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/testsuite/
* gcc.dg/pr96579.c: Compile only with target dfp.

Add expander for movp2hi and movp2qi.

2020-08-30  Uros Bizjak    <ubizjak@gmail.com>

gcc/ChangeLog:
PR target/96744
* config/i386/i386-expand.c (split_double_mode): Also handle
E_P2HImode and E_P2QImode.
* config/i386/sse.md (MASK_DWI): New define_mode_iterator.
(mov<mode>): New expander for P2HI,P2QI.
(*mov<mode>_internal): New define_insn_and_split to split
movement of P2QI/P2HI to 2 movqi/movhi patterns after reload.

gcc/testsuite/ChangeLog:

* gcc.target/i386/double_mask_reg-1.c: New test.

Fix: AVX512VP2INTERSECT should imply AVX512DQ.

gcc/ChangeLog

* common/config/i386/i386-common.c (ix86_handle_option): Set
AVX512DQ when AVX512VP2INTERSECT exists.

Daily bump.

gotest: use a space rather than a middle dot in target lists

Replace the U+00B7 middle dot character, placed after "mips64p32le"
in the target lists, with a space.  The U+00B7 character may not be
considered whitespace by Bourne shell and any non-ASCII character
may render incorrectly in some terminal devices.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/251177

libstdc++: Make std::chrono::duration use reduced ratio for period

This implements the changes from P0548 "common_type and duration". That
was a change for C++17, but as it corrects some issues introduced by DRs
I'm also treating it as a DR and changing it for all modes from C++11
up.

The main change is that duration<R,P>::period no longer denotes P, but
rather P::type, the reduced ratio. The unary operator+ and operator-
members of duration should now return a duration using that reduced
ratio.

The requirement that common_type<T>::type is the same type as
common_type<T, T>::type (rather than simply T) was already implemented
for PR 89102.

The standard says that duration::operator+() and duration::operator-()
should return common_type_t<duration>, but that seems unnecessarily
expensive to compute. This change just uses duration<rep, period> which
is the same type, so we don't need to instantiate common_type.

As an optimization, this also adds partial specializations of
common_type for two durations of the same type, a single duration, two
time_points of the same type, and a single time_point. These
specializations avoid instantiating other specializations of common_type
and one or both of __duration_common_type or __timepoint_common_type for
the cases where the answer is trivial to obtain.

libstdc++-v3/ChangeLog:

* include/std/chrono (__duration_common_type): Ensure the
reduced ratio is used. Remove unused partial specialization
using __failure_type.
(common_type): Pass reduced ratios to __duration_common_type.
Add partial specializations for simple cases involving a single
duration or time_point type.
(duration::period): Use reduced ratio.
(duration::operator+(), duration::operator-()): Return duration
type using the reduced ratio.
* testsuite/20_util/duration/requirements/typedefs_neg2.cc:
Adjust expected errors.
* testsuite/20_util/duration/requirements/reduced_period.cc: New test.

Fix shadd-2.c scan assembler count.

2020-08-27  John David Anglin  <danglin@gcc.gnu.org>

gcc/testsuite/
* gcc.target/hppa/shadd-2.c: Adjust times to 4.

libstdc++: Fix arithmetic bug in chrono::year_month::operator+

This fixes the months-based addition for year_month when the
year_month's month component is 0.

libstdc++-v3/ChangeLog:

* include/std/chrono (year_month::operator+): Properly handle a
month value of 0 by casting the month value to int before
subtracting 1 from it so that the difference is sign-extended in
the subsequent addition.
* testsuite/std/time/year_month/1.cc: Test adding months to a
year_month whose month component is below or above the
normalized range of [1,12].

libstdc++: Fix operator overload ambiguity with calendar types

We currently don't enforce a constraint on some of the calendar types'
addition/subtraction operator overloads that take a 'months' arguments:

  Constraints: If the argument supplied by the caller for the months
  parameter is convertible to years, its implicit conversion sequence to
  years is worse than its implicit conversion sequence to months.

This constraint is relevant when adding/subtracting a duration to/from,
say, a year_month where the given duration is convertible to both
'months' and to 'years' (as in the new testcases below).  The correct
behavior here in light of this constraint is to perform the operation
through the (more efficient) 'years'-based overload, but we currently
emit an ambiguous overload error.

This patch templatizes the 'months'-based addition/subtraction operator
overloads so that in the event of an implicit-conversion tie, we select
the non-template 'years'-based overload.  This is the same approach
that the date library takes for enforcing this constraint.

libstdc++-v3/ChangeLog:

* include/std/chrono
(__detail::__months_years_conversion_disambiguator): Define.
(year_month::operator+=): Templatize the 'months'-based overload
so that the 'years'-based overload is selected in case of
equally-ranked implicit conversion sequences to both 'months'
and 'years' from the supplied argument.
(year_month::operator-=): Likewise.
(year_month::operator+): Likewise.
(year_month::operator-): Likewise.
(year_month_day::operator+=): Likewise.
(year_month_day::operator-=): Likewise.
(year_month_day::operator+): Likewise.
(year_month_day::operator-): Likewise.
(year_month_day_last::operator+=): Likewise.
(year_month_day_last::operator-=): Likewise.
(year_month_day_last::operator+): Likewise
(year_month_day_last::operator-): Likewise.
(year_month_day_weekday::operator+=): Likewise
(year_month_day_weekday::operator-=): Likewise.
(year_month_day_weekday::operator+): Likewise.
(year_month_day_weekday::operator-): Likewise.
(year_month_day_weekday_last::operator+=): Likewise
(year_month_day_weekday_last::operator-=): Likewise.
(year_month_day_weekday_last::operator+): Likewise.
(year_month_day_weekday_last::operator-): Likewise.
(testsuite/std/time/year_month/2.cc): New test.
(testsuite/std/time/year_month_day/2.cc): New test.
(testsuite/std/time/year_month_day_last/2.cc): New test.
(testsuite/std/time/year_month_weekday/2.cc): New test.
(testsuite/std/time/year_month_weekday_last/2.cc): New test.

ia32: Fix alignment of _Atomic fields [PR65146]

For _Atomic fields, lowering the alignment of long long or double etc.
fields on ia32 is undesirable, because then one really can't perform atomic
operations on those using cmpxchg8b.

The following patch stops lowering the alignment in fields for _Atomic
types (the x86_field_alignment change) and for -mpreferred-stack-boundary=2
also ensures we don't misalign _Atomic long long etc. automatic variables
(the ix86_{local,minimum}_alignment changes).
Not sure about iamcu_alignment change, I know next to nothing about IA MCU,
but unless it doesn't have cmpxchg8b instruction, it would surprise me if we
don't want to do it as well.
clang apparently doesn't lower the field alignment for _Atomic.

2020-08-27  Jakub Jelinek  <jakub@redhat.com>

PR target/65146
* config/i386/i386.c (iamcu_alignment): Don't decrease alignment
for TYPE_ATOMIC types.
(ix86_local_alignment): Likewise.
(ix86_minimum_alignment): Likewise.
(x86_field_alignment): Likewise, and emit a -Wpsabi diagnostic
for it.

* gcc.target/i386/pr65146.c: New test.

rs6000: Support ELFv2 sibcall for indirect calls [PR96787]

Prior to P10, ELFv2 hasn't implemented nonlocal sibcalls.  Now that we do,
we need to be sure that r12 is set up prior to such a call.

2020-08-27  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
PR target/96787
* config/rs6000/rs6000.c (rs6000_sibcall_aix): Support
indirect call for ELFv2.

gcc/testsuite/

PR target/96787
* gcc.target/powerpc/pr96787-1.c: New.
* gcc.target/powerpc/pr96787-2.c: New.

Fortran  : ICE for division by zero in declaration PR95882

A length expression containing a divide by zero in a character
declaration will result in an ICE if the constant is anymore
complicated that a contant divided by a constant.

The cause was that char_len_param_value can return MATCH_YES
even if a divide by zero was seen.  Prior to returning check
whether a divide by zero was seen and if so set it to MATCH_ERROR.

2020-08-27  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/fortran

PR fortran/95882
* decl.c (char_len_param_value): Check gfc_seen_div0 and
if it is set return MATCH_ERROR.

2020-08-27  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite/

PR fortran/95882
* gfortran.dg/pr95882_1.f90: New test.
* gfortran.dg/pr95882_2.f90: New test.
* gfortran.dg/pr95882_3.f90: New test.
* gfortran.dg/pr95882_4.f90: New test.
* gfortran.dg/pr95882_5.f90: New test.

tree-optimization/96522 - transfer of flow-sensitive info in copy_ref_info

This removes the bogus tranfer of flow-sensitive info in copy_ref_info
plus fixes one oversight in FRE when flow-sensitive non-NULLness was added to
points-to info.

2020-08-27  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96522
* tree-ssa-address.c (copy_ref_info): Reset flow-sensitive
info of the copied points-to.  Transfer bigger alignment
via the access type.
* tree-ssa-sccvn.c (eliminate_dom_walker::eliminate_stmt):
Reset all flow-sensitive info.

* gcc.dg/torture/pr96522.c: New testcase.

vec: add exact argument for various grow functions.

gcc/ada/ChangeLog:

* gcc-interface/trans.c (gigi): Set exact argument of a vector
growth function to true.
(Attribute_to_gnu): Likewise.

gcc/ChangeLog:

* alias.c (init_alias_analysis): Set exact argument of a vector
growth function to true.
* calls.c (internal_arg_pointer_based_exp_scan): Likewise.
* cfgbuild.c (find_many_sub_basic_blocks): Likewise.
* cfgexpand.c (expand_asm_stmt): Likewise.
* cfgrtl.c (rtl_create_basic_block): Likewise.
* combine.c (combine_split_insns): Likewise.
(combine_instructions): Likewise.
* config/aarch64/aarch64-sve-builtins.cc (function_expander::add_output_operand): Likewise.
(function_expander::add_input_operand): Likewise.
(function_expander::add_integer_operand): Likewise.
(function_expander::add_address_operand): Likewise.
(function_expander::add_fixed_operand): Likewise.
* df-core.c (df_worklist_dataflow_doublequeue): Likewise.
* dwarf2cfi.c (update_row_reg_save): Likewise.
* early-remat.c (early_remat::init_block_info): Likewise.
(early_remat::finalize_candidate_indices): Likewise.
* except.c (sjlj_build_landing_pads): Likewise.
* final.c (compute_alignments): Likewise.
(grow_label_align): Likewise.
* function.c (temp_slots_at_level): Likewise.
* fwprop.c (build_single_def_use_links): Likewise.
(update_uses): Likewise.
* gcc.c (insert_wrapper): Likewise.
* genautomata.c (create_state_ainsn_table): Likewise.
(add_vect): Likewise.
(output_dead_lock_vect): Likewise.
* genmatch.c (capture_info::capture_info): Likewise.
(parser::finish_match_operand): Likewise.
* genrecog.c (optimize_subroutine_group): Likewise.
(merge_pattern_info::merge_pattern_info): Likewise.
(merge_into_decision): Likewise.
(print_subroutine_start): Likewise.
(main): Likewise.
* gimple-loop-versioning.cc (loop_versioning::loop_versioning): Likewise.
* gimple.c (gimple_set_bb): Likewise.
* graphite-isl-ast-to-gimple.c (translate_isl_ast_node_user): Likewise.
* haifa-sched.c (sched_extend_luids): Likewise.
(extend_h_i_d): Likewise.
* insn-addr.h (insn_addresses_new): Likewise.
* ipa-cp.c (gather_context_independent_values): Likewise.
(find_more_contexts_for_caller_subset): Likewise.
* ipa-devirt.c (final_warning_record::grow_type_warnings): Likewise.
(ipa_odr_read_section): Likewise.
* ipa-fnsummary.c (evaluate_properties_for_edge): Likewise.
(ipa_fn_summary_t::duplicate): Likewise.
(analyze_function_body): Likewise.
(ipa_merge_fn_summary_after_inlining): Likewise.
(read_ipa_call_summary): Likewise.
* ipa-icf.c (sem_function::bb_dict_test): Likewise.
* ipa-prop.c (ipa_alloc_node_params): Likewise.
(parm_bb_aa_status_for_bb): Likewise.
(ipa_compute_jump_functions_for_edge): Likewise.
(ipa_analyze_node): Likewise.
(update_jump_functions_after_inlining): Likewise.
(ipa_read_edge_info): Likewise.
(read_ipcp_transformation_info): Likewise.
(ipcp_transform_function): Likewise.
* ipa-reference.c (ipa_reference_write_optimization_summary): Likewise.
* ipa-split.c (execute_split_functions): Likewise.
* ira.c (find_moveable_pseudos): Likewise.
* lower-subreg.c (decompose_multiword_subregs): Likewise.
* lto-streamer-in.c (input_eh_regions): Likewise.
(input_cfg): Likewise.
(input_struct_function_base): Likewise.
(input_function): Likewise.
* modulo-sched.c (set_node_sched_params): Likewise.
(extend_node_sched_params): Likewise.
(schedule_reg_moves): Likewise.
* omp-general.c (omp_construct_simd_compare): Likewise.
* passes.c (pass_manager::create_pass_tab): Likewise.
(enable_disable_pass): Likewise.
* predict.c (determine_unlikely_bbs): Likewise.
* profile.c (compute_branch_probabilities): Likewise.
* read-rtl-function.c (function_reader::parse_block): Likewise.
* read-rtl.c (rtx_reader::read_rtx_code): Likewise.
* reg-stack.c (stack_regs_mentioned): Likewise.
* regrename.c (regrename_init): Likewise.
* rtlanal.c (T>::add_single_to_queue): Likewise.
* sched-deps.c (init_deps_data_vector): Likewise.
* sel-sched-ir.c (sel_extend_global_bb_info): Likewise.
(extend_region_bb_info): Likewise.
(extend_insn_data): Likewise.
* symtab.c (symtab_node::create_reference): Likewise.
* tracer.c (tail_duplicate): Likewise.
* trans-mem.c (tm_region_init): Likewise.
(get_bb_regions_instrumented): Likewise.
* tree-cfg.c (init_empty_tree_cfg_for_function): Likewise.
(build_gimple_cfg): Likewise.
(create_bb): Likewise.
(move_block_to_fn): Likewise.
* tree-complex.c (tree_lower_complex): Likewise.
* tree-if-conv.c (predicate_rhs_code): Likewise.
* tree-inline.c (copy_bb): Likewise.
* tree-into-ssa.c (get_ssa_name_ann): Likewise.
(mark_phi_for_rewrite): Likewise.
* tree-object-size.c (compute_builtin_object_size): Likewise.
(init_object_sizes): Likewise.
* tree-predcom.c (initialize_root_vars_store_elim_1): Likewise.
(initialize_root_vars_store_elim_2): Likewise.
(prepare_initializers_chain_store_elim): Likewise.
* tree-ssa-address.c (addr_for_mem_ref): Likewise.
(multiplier_allowed_in_address_p): Likewise.
* tree-ssa-coalesce.c (ssa_conflicts_new): Likewise.
* tree-ssa-forwprop.c (simplify_vector_constructor): Likewise.
* tree-ssa-loop-ivopts.c (addr_offset_valid_p): Likewise.
(get_address_cost_ainc): Likewise.
* tree-ssa-loop-niter.c (discover_iteration_bound_by_body_walk): Likewise.
* tree-ssa-pre.c (add_to_value): Likewise.
(phi_translate_1): Likewise.
(do_pre_regular_insertion): Likewise.
(do_pre_partial_partial_insertion): Likewise.
(init_pre): Likewise.
* tree-ssa-propagate.c (ssa_prop_init): Likewise.
(update_call_from_tree): Likewise.
* tree-ssa-reassoc.c (optimize_range_tests_cmp_bitwise): Likewise.
* tree-ssa-sccvn.c (vn_reference_lookup_3): Likewise.
(vn_reference_lookup_pieces): Likewise.
(eliminate_dom_walker::eliminate_push_avail): Likewise.
* tree-ssa-strlen.c (set_strinfo): Likewise.
(get_stridx_plus_constant): Likewise.
(zero_length_string): Likewise.
(find_equal_ptrs): Likewise.
(printf_strlen_execute): Likewise.
* tree-ssa-threadedge.c (set_ssa_name_value): Likewise.
* tree-ssanames.c (make_ssa_name_fn): Likewise.
* tree-streamer-in.c (streamer_read_tree_bitfields): Likewise.
* tree-vect-loop.c (vect_record_loop_mask): Likewise.
(vect_get_loop_mask): Likewise.
(vect_record_loop_len): Likewise.
(vect_get_loop_len): Likewise.
* tree-vect-patterns.c (vect_recog_mask_conversion_pattern): Likewise.
* tree-vect-slp.c (vect_slp_convert_to_external): Likewise.
(vect_bb_slp_scalar_cost): Likewise.
(vect_bb_vectorization_profitable_p): Likewise.
(vectorizable_slp_permutation): Likewise.
* tree-vect-stmts.c (vectorizable_call): Likewise.
(vectorizable_simd_clone_call): Likewise.
(scan_store_can_perm_p): Likewise.
(vectorizable_store): Likewise.
* expr.c: Likewise.
* vec.c (test_safe_grow_cleared): Likewise.
* vec.h (vec_safe_grow): Likewise.
(vec_safe_grow_cleared): Likewise.
(vl_ptr>::safe_grow): Likewise.
(vl_ptr>::safe_grow_cleared): Likewise.
* config/c6x/c6x.c (insn_set_clock): Likewise.

gcc/c/ChangeLog:

* gimple-parser.c (c_parser_gimple_compound_statement): Set exact argument of a vector
growth function to true.

gcc/cp/ChangeLog:

* class.c (build_vtbl_initializer): Set exact argument of a vector
growth function to true.
* constraint.cc (get_mapped_args): Likewise.
* decl.c (cp_maybe_mangle_decomp): Likewise.
(cp_finish_decomp): Likewise.
* parser.c (cp_parser_omp_for_loop): Likewise.
* pt.c (canonical_type_parameter): Likewise.
* rtti.c (get_pseudo_ti_init): Likewise.

gcc/fortran/ChangeLog:

* trans-openmp.c (gfc_trans_omp_do): Set exact argument of a vector
growth function to true.

gcc/lto/ChangeLog:

* lto-common.c (lto_file_finalize): Set exact argument of a vector
growth function to true.

streamline TARGET_MEM_REF dumping

The following streamlines TARGET_MEM_REF dumping building
on what we do for MEM_REF and thus dumping things like
access type, TBAA type and base/clique.  I've changed it
to do semantic dumping aka base + offset + step * index
rather than the odd base: A, step: way.

2020-08-27  Richard Biener  <rguenther@suse.de>

* tree-pretty-print.c (dump_mem_ref): Handle TARGET_MEM_REFs.
(dump_generic_node): Use dump_mem_ref also for TARGET_MEM_REF.

* gcc.dg/tree-ssa/loop-19.c: Adjust.
* gcc.dg/tree-ssa/loop-2.c: Likewise.
* gcc.dg/tree-ssa/loop-3.c: Likewise.

lra: Canonicalize mult to shift in address reloads

Inside a (mem) RTX, it is canonical to write multiplications by powers
of two using a (mult) [0]. Outside of a (mem), the canonical way to
write multiplications by powers of two is using (ashift).

Now I observed that LRA does not quite respect this RTL canonicalization
rule.  When compiling gcc/testsuite/gcc.dg/torture/pr34330.c with -Os
-ftree-vectorize, the RTL in the dump "281r.ira" has the insn:

(set (reg:SI 111)
     (mem:SI (plus:DI (mult:DI (reg:DI 101 [ ivtmp.9 ])
                 (const_int 4 [0x4]))
             (reg/v/f:DI 105 [ b ]))))

but LRA then proceeds to generate a reload, and we get the following
non-canonical insn in "282r.reload":

(set (reg:DI 7 x7 [121])
     (plus:DI (mult:DI (reg:DI 5 x5 [orig:101 ivtmp.9 ] [101])
             (const_int 4 [0x4]))
         (reg/v/f:DI 1 x1 [orig:105 b ] [105])))

This patch fixes LRA to ensure that we generate canonical RTL in this
case. After the patch, we get the following insn in "282r.reload":

(set (reg:DI 7 x7 [121])
        (plus:DI (ashift:DI (reg:DI 5 x5 [orig:101 ivtmp.9 ] [101])
                (const_int 2 [0x2]))
            (reg/v/f:DI 1 x1 [orig:105 b ] [105])))

[0] : https://gcc.gnu.org/onlinedocs/gccint/Insn-Canonicalizations.html

gcc/ChangeLog:

* lra-constraints.c (canonicalize_reload_addr): New.
(curr_insn_transform): Use canonicalize_reload_addr to ensure we
generate canonical RTL for an address reload.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/mem-shift-canonical.c: New test.

AArch64: Add FLAG for rounding intrinsics [PR94442]

2020-08-27  Zhiheng Xie  <xiezhiheng@huawei.com>

gcc/ChangeLog:

* config/aarch64/aarch64-simd-builtins.def: Add proper FLAG
for rounding intrinsics.

AArch64: Add FLAG for min/max intrinsics [PR94442]

2020-08-27  Zhiheng Xie  <xiezhiheng@huawei.com>

gcc/ChangeLog:

* config/aarch64/aarch64-simd-builtins.def: Add proper FLAG
for min/max intrinsics.

tree-optimization/96579 - another special-operands fix in reassoc

This makes sure to put special-ops expanded rhs left where
expression rewrite expects it.

2020-08-27  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96579
* tree-ssa-reassoc.c (linearize_expr_tree): If we expand
rhs via special ops make sure to swap operands.

* gcc.dg/pr96579.c: New testcase.

tree-optimization/96565 - improve DSE with paths ending in noreturn

This improves DSEs stmt walking by not considering a DEF without
uses for further processing (and thus giving up when there's two
paths to follow).

2020-08-26  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96565
* tree-ssa-dse.c (dse_classify_store): Remove defs with
no uses from further processing.

* gcc.dg/tree-ssa/ssa-dse-40.c: New testcase.
* gcc.dg/builtin-object-size-4.c: Adjust.

libstdc++: Implement remaining piece of LWG 3448

Almost all of the proposed resolution for LWG 3448 is already
implemented; the only part left is to adjust the return type of
transform_view::sentinel::operator-.

libstdc++-v3/ChangeLog:

PR libstdc++/95322
* include/std/ranges (transform_view::sentinel::__distance_from):
Give this a deduced return type.
(transform_view::sentinel::operator-): Adjust the return type so
that it's based on the constness of the iterator rather than
that of the sentinel.
* testsuite/std/ranges/adaptors/95322.cc: Refer to LWG 3488.

libstdc++: elements_view's sentinel and iterator not comparable [LWG 3406]

This implements the proposed resolution for LWG 3406, and adds a
testcase for the example from P1994R1.

libstdc++-v3/ChangeLog:

* include/std/ranges (elements_view::begin): Adjust constraints.
(elements_view::end): Likewise.
(elements_view::_Sentinel::operator==): Templatize to take both
_Iterator<true> and _Iterator<false>.
(elements_view::_Sentinel::operator-): Likewise.
* testsuite/std/ranges/adaptors/elements.cc: Add testcase for
the example from P1994R1.
* testsuite/std/ranges/adaptors/lwg3406.cc: New test.

libstdc++: Implement P1994R1 changes to ranges::elements_view

The example from the paper doesn't compile without the proposed
resolution for LWG 3406, so we'll add a testcase for this once the
proposed resolution is implemented.

libstdc++-v3/ChangeLog:

* include/std/ranges (elements_view::end): Replace these two
overloads with four new overloads.
(elements_view::_Iterator::operator==): Remove.
(elements_view::_Iterator::operator-): Likewise.
(elements_view::_Sentinel): Define.

Daily bump.

cmd: add -maix32 to gcc calls for aix/ppc

As gcc might now be compiled in 64bit, -maix32 must always be added
to ensure that created objects will be 32bit.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/250778

libgo: add FAT library support for static libraries on AIX

Like shared libraries, AIX static libraries must also have both 32 and
64 bit objects.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/250777

Mark various tests that require a c99 libm

A number of i386 math optimisation tests are looking assembly instructions
that are only emitted when the compiler knows the target has a C99 libm
available. Since targets like *-elf may not have such a libm, a C99 runtime
requirement is added to these tests.

gcc/testsuite/ChangeLog

* gcc.target/i386/387-7.c: Add dg-require-effective-target c99_runtime.
* gcc.target/i386/387-9.c: Likewise.
* g++.target/i386/avx512bw-pr96246-1.C: Likewise.
* gcc.target/i386/avx512f-rint-sfix-vec-2.c: Likewise.
* gcc.target/i386/avx512f-rintf-sfix-vec-2.c: Likewise.
* g++.target/i386/avx512vl-pr96246-1.C: Likewise.
* gcc.target/i386/pr61403.c: Likewise.
* gcc.target/i386/sse4_1-ceil-sfix-vec.c: Likewise.
* gcc.target/i386/sse4_1-ceilf-sfix-vec.c: Likewise.
* gcc.target/i386/sse4_1-floor-sfix-vec.c: Likewise.
* gcc.target/i386/sse4_1-floorf-sfix-vec.c: Likewise.
* gcc.target/i386/sse4_1-rint-sfix-vec.c: Likewise.
* gcc.target/i386/sse4_1-rintf-sfix-vec.c: Likewise.
* gcc.target/i386/sse4_1-round-sfix-vec.c: Likewise.
* gcc.target/i386/sse4_1-roundf-sfix-vec.c: Likewise.