Daily bump.

From-SVN: r269183

re PR c++/84585 (internal compiler error: in get_local_decls, at cp/name-lookup.c:3654)

2019-02-24  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/84585
* g++.dg/cpp0x/pr84585.C: New.

From-SVN: r269180

re PR fortran/89174 (Allocation segfault with CLASS(*) MOLD)

2019-02-24  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/89174
* trans-expr.c (gfc_find_and_cut_at_last_class_ref): Add is_mold
to garguments. If we are dealing with a MOLD, call
gfc_expr_to_initialize().
* trans-stmt.c (gfc_trans_allocate): For MOLD, pass is_mold=true
to gfc_find_and_cut_at_last_class_ref.
* trans.h (gfc_find_and_cut_at_last_class_ref): Add optional
argument is_mold with default false.

2019-02-24  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/89174
* gfortran.dg/allocate_with_mold_3.f90: New test.

From-SVN: r269179

i386: Compile PR target/87007 tests with -mfpmath=sse

-mfpmath=sse is needed to enable SSE for FP math in 32-bit.

PR target/87007
* gcc.target/i386/pr87007-1.c: Compile with -mfpmath=sse.
* gcc.target/i386/pr87007-2.c: Likewise.

From-SVN: r269178

re PR fortran/89266 (ICE with TRANSFER of len=0 character array constructor)

2019-02-24  Harald Anlauf  <anlauf@gmx.de>

PR fortran/89266
PR fortran/88326
* target-memory.c (gfc_element_size): Return false if element size
cannot be determined; element size is returned separately.
(gfc_target_expr_size): Return false if expression size cannot be
determined; expression size is returned separately.
* target-memory.h: Adjust prototypes.
* check.c (gfc_calculate_transfer_sizes): Adjust references to
gfc_target_expr_size, gfc_element_size.
* arith.c (hollerith2representation): Likewise.
* class.c (find_intrinsic_vtab): Likewise.
* simplify.c (gfc_simplify_sizeof): Likewise.

PR fortran/89266
PR fortran/88326
* gfortran.dg/pr89266.f90: New test.
* gfortran.dg/pr88326.f90: New test.

From-SVN: r269177

re PR rtl-optimization/89445 (_mm512_maskz_loadu_pd "forgets" to use the mask)

PR rtl-optimization/89445
* simplify-rtx.c (simplify_ternary_operation): Don't use
simplify_merge_mask on operands that may trap.
* rtlanal.c (may_trap_p_1): Use FLOAT_MODE_P instead of
SCALAR_FLOAT_MODE_P checks.  For integral division by zero, if
second operand is CONST_VECTOR, check if any element could be zero.
Don't expect traps for VEC_{MERGE,SELECT,CONCAT,DUPLICATE} unless
their operands can trap.

* gcc.target/i386/avx512f-pr89445.c: New test.

From-SVN: r269176

PR libstdc++/89416 fix accessibility of members

PR libstdc++/89416
* include/bits/alloc_traits.h (__is_alloc_insertable_impl): Make
copy and move members public.

From-SVN: r269175

Daily bump.

From-SVN: r269173

Improve error message for bad arguments to script

* generate_libstdcxx_web_docs: Improve error output.

From-SVN: r269169

P0340R2 Making std::underlying_type SFINAE-friendly

* include/std/type_traits (__underlying_type_impl): New helper to
make underlying_type SFINAE-friendly.
(underlying_type): Derive from __underlying_type_impl.
* testsuite/20_util/underlying_type/requirements/typedefs-3.cc: New
test.

From-SVN: r269168

builtin-sprintf-10.c: Cast remaining wchar_t to wint_t to avoid a bogus -Wformat warning in ILP32...

gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/builtin-sprintf-10.c: Cast remaining
wchar_t to wint_t to avoid a bogus -Wformat warning in
ILP32 (bug 77970).

From-SVN: r269167

gimple-ssa-sprintf.c (target_strtol): Rename...

gcc/ChangeLog:
* gimple-ssa-sprintf.c (target_strtol): Rename...
(target_strtohwi): ...to this.  Handle values up to HOST_WIDE_INT_MAX.
(parse_directive): Adjust to name change.  Use HOST_WIDE_INT_MAX to
check for range error.

From-SVN: r269166

Include "../../../libgfortran/ISO_Fortran_binding.h"

PR testsuite/89476
* gfortran.dg/ISO_Fortran_binding_5.c: Include
"../../../libgfortran/ISO_Fortran_binding.h".
* gfortran.dg/ISO_Fortran_binding_6.c: Likewise.

From-SVN: r269165

driver: Also prune joined switches with negation

When -march=native is passed to host_detect_local_cpu to the backend,
it overrides all command lines after it.  That means

$ gcc -march=native -march=skylake-avx512

is the treated as

$ gcc -march=skylake-avx512 -march=native

Prune joined switches with Negative and RejectNegative to allow
-march=skylake-avx512 to override previous -march=native on command-line.

gcc/

PR driver/69471
* opts-common.c (prune_options): Also prune joined switches
with Negative and RejectNegative.
* config/i386/i386.opt (march=): Add Negative(march=).
(mtune=): Add Negative(mtune=).
* doc/options.texi: Document Negative used together with Joined
and RejectNegative.

gcc/testsuite/

PR driver/69471
* gcc.dg/pr69471-1.c: New test.
* gcc.dg/pr69471-2.c: Likewise.
* gcc.target/i386/pr69471-3.c: Likewise.

From-SVN: r269164

re PR fortran/84387 (Defined output does not work for a derived type that has no components)

2019-02-23  Jerry DeLisle <jvdelisle@gcc.gnu.org>

PR fortran/84387
* trans-io.c (transfer_expr): Do not return if there are no
components to the derived type or class.

* gfortran.dg/dtio_34.f90: New test.

From-SVN: r269161

re PR c++/89419 (ICE in is_normal_capture_proxy starting with r253601)

PR c++/89419
* g++.dg/cpp1y/lambda-generic-89419.C: New test.

From-SVN: r269159

PR c++/88294 - ICE with non-constant noexcept-specifier.

* pt.c (maybe_instantiate_noexcept): Set up the list of local
specializations.  Set current_class_{ptr,ref}.

* g++.dg/cpp0x/noexcept34.C: New test.
* g++.dg/cpp0x/noexcept35.C: New test.

From-SVN: r269158

re PR fortran/88117 (ICE in gimplify_var_or_parm_decl, at gimplify.c:2697)

2019-02-23  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/88117
* resolve.c (deferred_op_assign): Return if the lhs expression
has the pointer attribute.
* trans-expr.c (gfc_trans_assignment_1): Do not fix the string
length if the lhs expression has the pointer attribute.

2019-02-23  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/88117
* gfortran.dg/deferred_character_32.f90 : New test

From-SVN: r269157

re PR fortran/89385 (Incorrect members of C descriptor for an allocatable object)

2019-02-23  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/89385
PR fortran/89366
* decl.c (gfc_verify_c_interop_param): Restriction on string
length being one is lifted for F2018.
* trans-expr.c (gfc_conv_gfc_desc_to_cfi_desc): For scalar
characters with intent in, make a temporary and copy the result
of the expression evaluation into it.
(gfc_conv_procedure_call): Set a flag for character formal args
having a character length that is not unity. If the procedure
is bind C, call gfc_conv_gfc_desc_to_cfi_desc in this case.
Also, extend bind C calls to unconditionally convert both
pointers and allocatable expressions.

2019-02-23  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/89385
* gfortran.dg/ISO_Fortran_binding_1.f90 : Correct test for
previously incorrect lbound for allocatable expressions. Also
correct stop values to avoid repetition.
* gfortran.dg/ISO_Fortran_binding_5.f90 : New test
* gfortran.dg/ISO_Fortran_binding_5.c : Support previous test.

PR fortran/89366
* gfortran.dg/ISO_Fortran_binding_6.f90 : New test
* gfortran.dg/ISO_Fortran_binding_6.c : Support previous test.
* gfortran.dg/pr32599.f03 : Set standard to F2008.

2019-02-23  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/89385
PR fortran/89366
* runtime/ISO_Fortran_binding.c (cfi_desc_to_gfc_desc) : In the
interchange between character and derived, the character type
was being set incorrectly.
(gfc_desc_to_cfi_desc) : Eliminate the interchange of types in
this function. Do not add the kind and length information to
the type field of structures. Lbounds were incorrectly being
set to zero for allocatable and pointer descriptors. Should
have been non-pointer, non-allocatables that received this
treatment.

From-SVN: r269156

PR libstdc++/89446 fix null pointer dereference in char_traits

PR libstdc++/89446
* include/bits/char_traits.h (__constant_char_array): Check index is
in range before dereferencing.
(char_traits<char>::compare, char_traits<char>::find)
(char_traits<char8_t>::compare, char_traits<char8_t>::find): Return
immediately if n is zero.
(char_traits<wchar_t>::compare, char_traits<wchar_t>::find): Likewise.
Remove workarounds for PR 67026.
* testsuite/21_strings/basic_string_view/operators/char/89446.cc:
New test.
* testsuite/21_strings/basic_string_view/operators/wchar_t/89446.cc:
New test.

From-SVN: r269148

Capture source location of dtors (PR c++/89390)

gcc/cp/ChangeLog:
PR c++/89390
* parser.c (cp_parser_unqualified_id): Capture and use locations
for destructors.

gcc/testsuite/ChangeLog:
PR c++/89390
* g++.dg/diagnostic/pr89390.C: Update expected location of error,
renaming to a multicharacter name, so that start != finish.  Add
tests for dtor locations.

From-SVN: r269145

Daily bump.

From-SVN: r269142

re PR tree-optimization/88074 (g++ hangs on math expression)

PR middle-end/88074
* simplify.c (norm2_do_sqrt, gfc_simplify_norm2): Use
mpfr_number_p && !mpfr_zero_p instead of mpfr_regular_p.
(norm2_add_squared): Likewise.  Use mp_exp_t rather than mpfr_exp_t.

Co-Authored-By: Jakub Jelinek <jakub@redhat.com>
From-SVN: r269139

re PR c++/84676 (internal compiler error: Segmentation fault (build_new_op_1))

2019-02-22  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/84676
* g++.dg/cpp0x/pr84676.C: New.

From-SVN: r269138

builtin-sprintf-10.c: Cast wchar_t to wint_t to avoid a bogus -Wformat warning in ILP32 (bug 77970).

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/builtin-sprintf-10.c: Cast wchar_t to wint_t
to avoid a bogus -Wformat warning in ILP32 (bug 77970).

From-SVN: r269137

baseline_symbols.txt: Adjust.

* config/abi/post/sparc64-linux-gnu/baseline_symbols.txt: Adjust.
* config/abi/post/sparc64-linux-gnu/32/baseline_symbols.txt: Likewise.

From-SVN: r269136

re PR fortran/83057 (OPEN without a filename and without STATUS='SCRATCH' could produce a warning)

2019-02-22  Harald Anlauf  <anlauf@gmx.de>

PR fortran/83057
* io.c (gfc_match_open): Fix logic in checks of OPEN statement
when NEWUNIT= is specified.

PR fortran/83057
* gfortran.dg/newunit_6.f90: New test.

From-SVN: r269134

re PR fortran/89431 (CPP integer macros not defined)

2019-02-22  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/89431
* gfortran.texi: Fix documentation to match the implementation.

From-SVN: r269132

PR c++/89420 - ICE with CAST_EXPR in explicit-specifier.

* decl.c (build_explicit_specifier): Don't check
processing_template_decl.  Call instantiation_dependent_expression_p
instead of value_dependent_expression_p.  Call
instantiate_non_dependent_expr_sfinae before
build_converted_constant_expr instead of calling
instantiate_non_dependent_expr after it.  Add
processing_template_decl_sentinel.

* g++.dg/cpp2a/explicit14.C: New test.

From-SVN: r269131

re PR libstdc++/89402 (warning: ‘void _ZNKSt4hashIeEclEe()’ specifies less restrictive attribute than its target)

PR libstdc++/89402
* src/c++98/compatibility-ldbl.cc (_ZNKSt4hashIeEclEe): Change return
type to std::size_t and argument to type to long double.

From-SVN: r269130

extend.texi (Other Builtins): Add __builtin_is_constant_evaluated.

gcc/ChangeLog:

* doc/extend.texi (Other Builtins): Add
__builtin_is_constant_evaluated.

From-SVN: r269129

re PR tree-optimization/87609 (miscompilation with restrict and loop)

2019-02-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87609
* tree-cfg.c (gimple_duplicate_bb): Only remap inlined cliques.

From-SVN: r269127

PR tree-optimization/88993 - GCC 9 -Wformat-overflow=2 should reflect real libc limits

PR tree-optimization/88993 - GCC 9 -Wformat-overflow=2 should reflect real libc limits
PR tree-optimization/88835 - overly aggressive -Werror=format-overflow for printf

gcc/ChangeLog:

PR tree-optimization/88993
PR tree-optimization/88853
* gimple-ssa-sprintf.c (sprintf_dom_walker::call_info::is_file_func):
New helper.
(sprintf_dom_walker::call_info::is_string_func): New helper.
(format_directive): Only issue "may exceed" 4095/INT_MAX warnings
for formatted string functions.
(sprintf_dom_walker::handle_gimple_call): Fix a typo in a comment.

gcc/testsuite/ChangeLog:

PR tree-optimization/88993
PR tree-optimization/88853
* gcc.dg/tree-ssa/builtin-fprintf-warn-2.c: New test.
* gcc.dg/tree-ssa/builtin-printf-warn-2.c: New test.
* gcc.dg/tree-ssa/builtin-snprintf-warn-3.c: Adjust.
* gcc.dg/tree-ssa/builtin-sprintf-warn-18.c: Same.

From-SVN: r269125

re PR rtl-optimization/87761 ([MIPS] New FAIL: gcc.target/mips/fix-r4000-10.c   -O1 start with r265398)

PR rtl-optimization/87761
* config/mips/mips.md: Add new combiner pattern to recognize
a bitfield extraction using (ashiftrt (truncate (ashift (...)))).

From-SVN: r269123

Handle stack pointer with SUBS/ADDS instructions.

In general the stack pointer was not handled for many SUBS/ADDS patterns in
aarch64.md.
Both the "extended register" and "immediate" forms allow the stack pointer to be
used as the source register, while no form allows the stack pointer for the
destination register.

The define_insn patterns generating ADDS/SUBS did not allow the stack pointer
for any operand, while the define_peephole2 patterns that generated RTX to be
matched by these patterns allowed the stack pointer for any operand.

The patterns are fixed by adding the 'k' constraint for the first source operand
to all define_insns that generate the ADDS/SUBS "extended register" and
"immediate" forms (but not the "shifted register" form).

In peephole optimizations, constraint strings are ignored (see "(gccint) C
Constraint Interface" info node in the documentation), so the decision to act or
not is based solely on the predicate and condition.
This patch introduces a new predicate "aarch64_general_reg" to be used in
define_peephole2 patterns where only GENERAL_REGS registers are acceptable and
uses that predicate in the peepholes that generate patterns for ADDS/SUBS.

Full bootstrap and regtest done on aarch64-none-linux-gnu.
Regression tests done on aarch64-none-linux-gnu and aarch64-none-elf cross
compiler.

OK for trunk?

gcc/ChangeLog:

2019-02-22  Matthew Malcomson  <matthew.malcomson@arm.com>

PR target/89324
* config/aarch64/aarch64.md: Use aarch64_general_reg predicate on
destination register in peepholes generating patterns for ADDS/SUBS.
(add<mode>3_compare0,
*addsi3_compare0_uxtw, add<mode>3_compareC,
add<mode>3_compareV_imm, add<mode>3_compareV,
*adds_<optab><ALLX:mode>_<GPI:mode>,
*subs_<optab><ALLX:mode>_<GPI:mode>,
*adds_<optab><ALLX:mode>_shift_<GPI:mode>,
*subs_<optab><ALLX:mode>_shift_<GPI:mode>,
*adds_<optab><mode>_multp2, *subs_<optab><mode>_multp2,
*sub<mode>3_compare0, *subsi3_compare0_uxtw,
sub<mode>3_compare1): Allow stack pointer for source register.
* config/aarch64/predicates.md (aarch64_general_reg): New predicate.

gcc/testsuite/ChangeLog:

2019-02-22  Matthew Malcomson  <matthew.malcomson@arm.com>

PR target/89324
* gcc.dg/rtl/aarch64/subs_adds_sp.c: New test.
* gfortran.fortran-torture/compile/pr89324.f90: New test.

From-SVN: r269122

PR c/89425 - -Wabsolute-value warns in dead subexpressions

gcc/c/ChangeLog:

PR c/89425
* c-parser.c (sizeof_ptr_memacc_comptypes): Avoid warning in
unreachable subexpressions.

gcc/testsuite/ChangeLog:

PR c/89425
* gcc.dg/Wabsolute-value.c: New test.

From-SVN: r269121

Wbuiltin-declaration-mismatch-12.c: New test.

gcc/testsuite/ChangeLog:
* gcc.dg/Wbuiltin-declaration-mismatch-12.c: New test.

From-SVN: r269120

i386: Add pass_remove_partial_avx_dependency

With -mavx, for

$ cat foo.i
extern float f;
extern double d;
extern int i;

void
foo (void)
{
  d = f;
  f = i;
}

we need to generate

vxorp[ds] %xmmN, %xmmN, %xmmN
...
vcvtss2sd f(%rip), %xmmN, %xmmX
...
vcvtsi2ss i(%rip), %xmmN, %xmmY

to avoid partial XMM register stall.  This patch adds a pass to generate
a single

vxorps %xmmN, %xmmN, %xmmN

at entry of the nearest dominator for basic blocks with SF/DF conversions,
which is in the fake loop that contains the whole function, instead of
generating one

vxorp[ds] %xmmN, %xmmN, %xmmN

for each SF/DF conversion.

NB: The LCM algorithm isn't appropriate here since it may place a vxorps
inside the loop.  Simple testcase show this:

$ cat badcase.c

extern float f;
extern double d;

void
foo (int n, int k)
{
  for (int j = 0; j != n; j++)
    if (j < k)
      d = f;
}

It generates

    ...
    loop:
      if(j < k)
        vxorps    %xmm0, %xmm0, %xmm0
        vcvtss2sd f(%rip), %xmm0, %xmm0
      ...
    loopend
    ...

This is because LCM only works when there is a certain benifit.  But for
conditional branch, LCM wouldn't move

   vxorps  %xmm0, %xmm0, %xmm0

out of loop.  SPEC CPU 2017 on Intel Xeon with AVX512 shows:

1. The nearest dominator

|RATE |Improvement|
|500.perlbench_r | 0.55% |
|538.imagick_r | 8.43% |
|544.nab_r | 0.71% |

2. LCM

|RATE |Improvement|
|500.perlbench_r | -0.76% |
|538.imagick_r | 7.96%  |
|544.nab_r | -0.13% |

Performance impacts of SPEC CPU 2017 rate on Intel Xeon with AVX512
using

-Ofast -flto -march=skylake-avx512 -funroll-loops

before

commit e739972ad6ad05e32a1dd5c29c0b950a4c4bd576
Author: uros <uros@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Thu Jan 31 20:06:42 2019 +0000

            PR target/89071
            * config/i386/i386.md (*extendsfdf2): Split out reg->reg
            alternative to avoid partial SSE register stall for TARGET_AVX.
            (truncdfsf2): Ditto.
            (sse4_1_round<mode>2): Ditto.

    git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@268427 138bc75d-0d04-0410-961f-82ee72b054a4

are:

|INT RATE |Improvement|
|500.perlbench_r | 0.55% |
|502.gcc_r | 0.14% |
|505.mcf_r | 0.08% |
|523.xalancbmk_r | 0.18% |
|525.x264_r |-0.49% |
|531.deepsjeng_r |-0.04% |
|541.leela_r |-0.26% |
|548.exchange2_r |-0.3% |
|557.xz_r |BuildSame|

|FP RATE |Improvement|
|503.bwaves_r         |-0.29% |
|507.cactuBSSN_r | 0.04% |
|508.namd_r |-0.74% |
|510.parest_r |-0.01% |
|511.povray_r | 2.23% |
|519.lbm_r | 0.1% |
|521.wrf_r | 0.49% |
|526.blender_r | 0.13% |
|527.cam4_r | 0.65% |
|538.imagick_r | 8.43% |
|544.nab_r | 0.71% |
|549.fotonik3d_r | 0.15% |
|554.roms_r | 0.08% |

After commit e739972ad6ad05e32a1dd5c29c0b950a4c4bd576, on Skylake client,
impacts on 538.imagick_r with

-fno-unsafe-math-optimizations -march=native -Ofast -funroll-loops -flto

1. Size comparision:

before:

   text    data     bss     dec     hex filename
2436377    8352    4528 2449257  255f69 imagick_r

after:

   text    data     bss     dec     hex filename
2425249    8352    4528 2438129  2533f1 imagick_r

2. Number of vxorps:

before after difference
4948            4135            -19.66%

3. Performance improvement:

|RATE |Improvement|
|538.imagick_r | 5.5%  |

gcc/

2019-02-22  H.J. Lu  <hongjiu.lu@intel.com>
    Hongtao Liu  <hongtao.liu@intel.com>
    Sunil K Pandey  <sunil.k.pandey@intel.com>

PR target/87007
* config/i386/i386-passes.def: Add
pass_remove_partial_avx_dependency.
* config/i386/i386-protos.h
(make_pass_remove_partial_avx_dependency): New.
* config/i386/i386.c (make_pass_remove_partial_avx_dependency):
New function.
(pass_data_remove_partial_avx_dependency): New.
(pass_remove_partial_avx_dependency): Likewise.
(make_pass_remove_partial_avx_dependency): Likewise.
* config/i386/i386.md (avx_partial_xmm_update): New attribute.
(*extendsfdf2): Add avx_partial_xmm_update.
(truncdfsf2): Likewise.
(*float<SWI48:mode><MODEF:mode>2): Likewise.
(SF/DF conversion splitters): Disabled for TARGET_AVX.

gcc/testsuite/

2019-02-22  H.J. Lu  <hongjiu.lu@intel.com>
    Hongtao Liu  <hongtao.liu@intel.com>
    Sunil K Pandey  <sunil.k.pandey@intel.com>

PR target/87007
* gcc.target/i386/pr87007-1.c: New test.
* gcc.target/i386/pr87007-2.c: Likewise.

Co-Authored-By: Hongtao Liu <hongtao.liu@intel.com>
Co-Authored-By: Sunil K Pandey <sunil.k.pandey@intel.com>
From-SVN: r269119

re PR middle-end/85598 (Incorrect -Wformat-truncation in a loop only at -O2 and -O3)

PR middle-end/85598
* gimple-ssa-sprintf.c (pass_sprintf_length::execute): Enable loop
analysis for pass.

From-SVN: r269115

x86: (Reapply) Move AESNI generation to Skylake and Goldmont

This is a repeat of commit r263989, which commit r264052 accidentally
reverted.

2019-02-22  Thiago Macieira  <thiago.macieira@intel.com>

PR target/89444
* config/i386/i386.h (PTA_WESTMERE): Remove PTA_AES.
(PTA_SKYLAKE): Add PTA_AES.
(PTA_GOLDMONT): Likewise.

From-SVN: r269114

c-pch.c (no_checksum): Remove.

2019-02-22  Richard Biener  <rguenther@suse.de>

c-family/
* c-pch.c (no_checksum): Remove.
(pch_init): Remove assertion that executable_checksum is not
all zero.
(c_common_valid_pch): Likewise.

From-SVN: r269113

[GCC, AArch64] Fix a couple of bugs in BTI

While doing more testing I found a couple of issues with my BTI patches.
This patch fixes them:
1) Remove a reference to return address key. The original patch was written
based on a different not yet committed patch ([PATCH 3/3][GCC][AARCH64] Add
support for pointer authentication B key) and I missed out on cleaning this
up. This is hidden behind the configuration option and thus went unnoticed.
2) Add a missed case for adding the BTI instruction in thunk functions.

*** gcc/ChangeLog ***

2019-02-22  Sudakshina Das  <sudi.das@arm.com>

* config/aarch64/aarch64.c (aarch64_output_mi_thunk): Add bti
instruction if enabled.
(aarch64_override_options): Remove reference to return address key.

From-SVN: r269112

re PR tree-optimization/89440 (ICE in vect_create_epilog_for_reduction, at tree-vect-loop.c:5502)

2019-02-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/89440
* tree-vect-loop.c (vect_create_epilog_for_reduction): Remove
not necessary assert.

* gcc.dg/vect/pr89440.c: New testcase.

From-SVN: r269111

Remove unused check_effective_target_offload_nvptx

gcc/testsuite/
* lib/target-supports.exp (check_effective_target_offload_nvptx):
Remove.

From-SVN: r269110

[libgomp] In OpenACC testing, by default only build for the offload target that we're actually going to test

... to avoid compilation overhead, and to keep simple '-foffload=[...]'
handling in test cases.

libgomp/
* testsuite/libgomp.oacc-c++/c++.exp: Specify
"-foffload=$offload_target".
* testsuite/libgomp.oacc-c/c.exp: Likewise.
* testsuite/libgomp.oacc-fortran/fortran.exp: Likewise.
* testsuite/lib/libgomp.exp
(check_effective_target_openacc_nvidia_accel_configured): Remove,
as (conceptually) merged into
check_effective_target_openacc_nvidia_accel_selected.  Adjust all
users.

From-SVN: r269109

[libgomp] In OpenACC testing, cycle though all offload targets

... instead of through offload plugins.

libgomp/
* plugin/configfrag.ac: Populate and AC_SUBST offload_targets.
* testsuite/libgomp-test-support.exp.in: Adjust.
* testsuite/lib/libgomp.exp: Likewise.  Don't populate
openacc_device_types_s.
(offload_target_to_openacc_device_type): New proc.
* testsuite/libgomp.oacc-c++/c++.exp: Adjust.
* testsuite/libgomp.oacc-c/c.exp: Likewise.
* testsuite/libgomp.oacc-fortran/fortran.exp: Likewise.
* Makefile.in: Regenerate.
* configure: Likewise.
* testsuite/Makefile.in: Likewise.

From-SVN: r269108

[libgomp] Clarify difference between offload target, offload plugin, and OpenACC device type

libgomp/
* plugin/configfrag.ac: Populate and AC_SUBST offload_plugins
instead of offload_targets, and AC_DEFINE_UNQUOTED OFFLOAD_PLUGINS
instead of OFFLOAD_TARGETS.
* target.c (gomp_target_init): Adjust.
* testsuite/libgomp-test-support.exp.in: Likewise.
* testsuite/lib/libgomp.exp: Likewise.  Populate
openacc_device_types_s instead of offload_targets_s_openacc.
(check_effective_target_openacc_nvidia_accel_selected)
(check_effective_target_openacc_host_selected): Adjust.
* testsuite/libgomp.oacc-c++/c++.exp: Likewise.
* testsuite/libgomp.oacc-c/c.exp: Likewise.
* testsuite/libgomp.oacc-fortran/fortran.exp: Likewise.
* Makefile.in: Regenerate.
* config.h.in: Likewise.
* configure: Likewise.
* testsuite/Makefile.in: Likewise.

From-SVN: r269107

[libgomp] In OpenACC offloading testing, be more explicit in what is supported, and what is not, or why not

libgomp/
* testsuite/lib/libgomp.exp: Error out for unknown offload target.
* testsuite/libgomp.oacc-c++/c++.exp: Likewise.  Report if
"offloading: supported, but hardware not accessible".
* testsuite/libgomp.oacc-c/c.exp: Likewise.
* testsuite/libgomp.oacc-fortran/fortran.exp: Likewise.

From-SVN: r269106

[PR72741] Use 'oacc_build_routine_dims' for Fortran OpenACC 'routine' directives, too

... instead of having an incomplete local implementation.

With these changes in place, we can then also revert the work-around r267213
"[nvptx] Unify C/Fortran routine handling in nvptx_goacc_validate_dims".

gcc/fortran/
PR fortran/72741
* gfortran.h (oacc_routine_lop): New enum.
(symbol_attribute): Use it.
* openmp.c (gfc_oacc_routine_dims): Replace with...
(gfc_oacc_routine_lop): ... this new function.
(gfc_match_oacc_routine): Adjust.
* trans-decl.c (add_attributes_to_decl): Likewise.
gcc/
PR fortran/72741
* omp-general.c (oacc_replace_fn_attrib): Mostly split out into...
(oacc_replace_fn_attrib_attr): ... this new function.
* omp-general.h (oacc_replace_fn_attrib_attr): New prototype.
* config/nvptx/nvptx.c (nvptx_goacc_validate_dims_1): Revert workaround.
gcc/testsuite/
PR fortran/72741
* gfortran.dg/goacc/classify-routine.f95: Adjust.

Co-Authored-By: Cesar Philippidis <cesar@codesourcery.com>
From-SVN: r269105

Revert earlier OpenACC 'routine' changes

..., which were committed as part of, but logically don't belong into r261813
"Update OpenACC data clause semantics to the 2.5 behavior", and which confuse
later changes.

gcc/fortran/
* openmp.c (gfc_match_oacc_declare): Revert earlier changes.
gcc/testsuite/
* c-c++-common/goacc/routine-5.c: Revert earlier changes.
* g++.dg/goacc/template.C: Likewise.

From-SVN: r269104

Silence '-Whsa' diagnostic in 'gfortran.dg/goacc/pr78027.f90'

... which has been present (with HSA offloading configured) ever since this
test case got added.

gcc/testsuite/
PR fortran/78027
* gfortran.dg/goacc/pr78027.f90: Add 'dg-additional-options "-Wno-hsa"'.

From-SVN: r269103

[C, C++] Use correct location information for OpenACC shape and simple clauses

gcc/c/
* c-parser.c (c_parser_oacc_shape_clause): Add loc formal
parameter.  Adjust all users.
(c_parser_oacc_simple_clause): Replace parser with loc formal
parameter.  Adjust all users.
gcc/cp/
* parser.c (cp_parser_oacc_simple_clause): Remove parser formal
parameter, move loc formal parameter to the front.  Adjust all
users.
(cp_parser_oacc_shape_clause): Add loc formal parameter.  Adjust
all users.

From-SVN: r269102

[arm] Add support for Neoverse N1

This patch adds support for the Neoverse N1 [1]. This CPU was previously supported through the Ares codename.
-mcpu=ares is retained as an alias of the new -mcpu=neoverse-n1.

Bootstrapped and tested on arm-none-linux-gnueabihf.

* config/arm/arm-cpus.in (ares): Rename to...
(neoverse-n1): ... This.  Add ares as alias.
* config/arm/arm-tables.opt: Regenerate.
* config/arm/arm-tune.md: Likewise.
* doc/invoke.txt (ARM Options): Document neoverse-n1.

From-SVN: r269101

[AArch64] Add support for Neoverse E1

This patch adds -mcpu and -mtune support for the Neoverse E1 CPU [1].
The new option is -mcpu=neoverse-e1.
Bootstrapped and tested on aarch64-none-linux-gnu.

* config/aarch64/aarch64-cores.def (neoverse-e1): Define.
* config/aarch64/aarch64-tune.md: Regenerate.
* doc/invoke.texi (AArch64 Options): Document neoverse-e1 -mcpu option.

From-SVN: r269100

[AArch64] Add support for Neoverse N1

This patch adds support for the Neoverse N1 CPU [1]. This was supported in GCC earlier through the codename Ares,
which it now replaces. -mcpu=ares is still accepted as there's been a binutils release supporting it,
but the internal structures are renamed to use Neoverse N1-related identifiers.

Bootstrapped and tested on aarch64-none-linux-gnu.

* config/aarch64/aarch64.c (ares_tunings): Rename to...
(neoversen1_tunings): ... This.
* config/aarch64/aarch64-cores.def (ares): Change tuning to the above.
(neoverse-n1): New CPU.
* config/aarch64/aarch64-tune.md: Regenerate.
* doc/invoke.txt (AArch64 Options): Document neoverse-n1.

From-SVN: r269099

re PR tree-optimization/87609 (miscompilation with restrict and loop)

2019-02-22  Richard Biener  <rguenther@suse.de>

PR middle-end/87609
* cfghooks.h (dependence_hash): New typedef.
(struct copy_bb_data): New type.
(cfg_hooks::duplicate_block): Adjust to take a copy_bb_data argument.
(duplicate_block): Likewise.
* cfghooks.c (duplicate_block): Pass down copy_bb_data.
(copy_bbs): Create and pass down copy_bb_data.
* cfgrtl.c (cfg_layout_duplicate_bb): Adjust.
(rtl_duplicate_bb): Likewise.
* tree-cfg.c (gimple_duplicate_bb): If the copy_bb_data arg is not NULL
remap dependence info.

* gcc.dg/torture/restrict-7.c: New testcase.

From-SVN: r269098

re PR tree-optimization/87609 (miscompilation with restrict and loop)

2019-02-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87609
* tree-core.h (tree_base): Document special clique values.
* tree-inline.c (remap_dependence_clique): Do not use the
special clique value of one.
(maybe_set_dependence_info): Use clique one.
(clear_dependence_clique): New callback.
(compute_dependence_clique): Clear clique one from all refs
before assigning it (again).

From-SVN: r269097

configure.host (abi_baseline_pair): Adjust for SPARC64/Linux.

* configure.host (abi_baseline_pair): Adjust for SPARC64/Linux.
* config/abi/post/sparc64-linux-gnu: New directory.
* config/abi/post/sparc64-linux-gnu/baseline_symbols.txt: New file.
* config/abi/post/sparc64-linux-gnu/32: New directory.
* config/abi/post/sparc64-linux-gnu/32/baseline_symbols.txt: New file.

From-SVN: r269096

PR c++/87685 - generic lambda 'this' capture error.

The standard says that in a generic lambda we should speculatively capture
'this' if we see a call to an overload set that contains a non-static member
function, but it seems wrong to reject the program if we can't capture,
since it might not actually be needed.

* lambda.c (lambda_expr_this_capture): Change add_capture_p to int.
(maybe_generic_this_capture): Pass -1.

From-SVN: r269095

PR c++/88394 - ICE with VLA init-capture.

We mostly use is_normal_capture_proxy to decide whether or not to use
DECL_CAPTURED_VARIABLE; we could just check whether it's set.  VLA capture
is still mostly broken, but this fixes this ICE.

* lambda.c (is_normal_capture_proxy): Check DECL_CAPTURED_VARIABLE.

From-SVN: r269094

PR c++/88869 - C++17 ICE with CTAD and explicit specialization.

The members of an explicit specialization of a class template don't have the
template parameters of that class template, so we shouldn't try to provide
arguments for them.  Only set outer_args when the class is an instantiation.

* pt.c (do_class_deduction): Don't include explicit specialization
args in outer_args.

From-SVN: r269093

Add tests for C++2a content of <atomic> header

* testsuite/29_atomics/headers/atomic/types_std_c++20.cc: New test.
* testsuite/29_atomics/headers/atomic/types_std_c++20_neg.cc: New
test.

From-SVN: r269091

P0482R5 char8_t: Updates to gdb pretty printing support

This patch adds recognition of the u8string and u8string_view type
aliases to the gdb pretty printer extension.

2019-02-22  Tom Honermann  <tom@honermann.net>

* python/libstdcxx/v6/printers.py (register_type_printers): Add type
printers for u8string and u8string_view.

From-SVN: r269090

P0482R5 char8_t: Updates to existing standard library tests

This patch augments existing tests to validate behavior for char8_t.  In
all cases, added test cases are cloned from existing tests for wchar_t
or char16_t.

A few tests required updates to line numbers for diagnostic messages.

2019-02-22  Tom Honermann  <tom@honermann.net>

* testsuite/18_support/byte/ops.cc: Validate
std::to_integer<char8_t>, std::to_integer<char16_t>, and
std::to_integer<char32_t>.
* testsuite/18_support/numeric_limits/dr559.cc: Validate
std::numeric_limits<char8_t>.
* testsuite/18_support/numeric_limits/lowest.cc: Validate
std::numeric_limits<char8_t>::lowest().
* testsuite/18_support/numeric_limits/max_digits10.cc: Validate
std::numeric_limits<char8_t>::max_digits10.
* testsuite/18_support/type_info/fundamental.cc: Validate
typeinfo for char8_t.
* testsuite/20_util/from_chars/1_c++20_neg.cc: New test, validating
std::from_chars with char8_t.
* testsuite/20_util/hash/requirements/explicit_instantiation.cc:
Validate explicit instantiation of std::hash<char8_t>.
* testsuite/20_util/is_integral/value.cc: Validate
std::is_integral<char8_t>.
* testsuite/20_util/make_signed/requirements/typedefs-4.cc:
Validate std::make_signed<char8_t>.
* testsuite/21_strings/basic_string/cons/char/deduction.cc:
Validate u8string construction from char8_t sources.
* testsuite/21_strings/basic_string/types/pmr_typedefs.cc: Validate
std::pmr::u8string.
* testsuite/21_strings/basic_string_view/operations/compare/
char/70483.cc: Validate substr operations on u8string_view.
* testsuite/21_strings/basic_string_view/typedefs.cc: Validate that
the u8string_view typedef is defined.
* testsuite/21_strings/char_traits/requirements/
constexpr_functions.cc: Validate char_traits<char8_t> constexpr
member functions.
* testsuite/21_strings/char_traits/requirements/
constexpr_functions_c++17.cc: Validate char_traits<char8_t> C++17
constexpr member functions.
* testsuite/21_strings/headers/string/types_std_c++0x.cc: Validate
that the u8string typedef is defined.
* testsuite/22_locale/locale/cons/unicode.cc: Validate the presence
of the std::codecvt<char16_t, char8_t, std::mbstate_t> and
std::codecvt<char32_t, char8_t, std::mbstate_t> facets.
* testsuite/29_atomics/atomic/cons/assign_neg.cc: Update line
numbers.
* testsuite/29_atomics/atomic/cons/copy_neg.cc: Likewise.
* testsuite/29_atomics/atomic_integral/cons/assign_neg.cc:
Likewise.
* testsuite/29_atomics/atomic_integral/cons/copy_neg.cc: Likewise.
* testsuite/29_atomics/atomic_integral/is_always_lock_free.cc:
Validate std::atomic<char8_t>::is_always_lock_free
* testsuite/29_atomics/atomic_integral/operators/bitwise_neg.cc:
Update line numbers.
* testsuite/29_atomics/atomic_integral/operators/decrement_neg.cc:
Likewise.
* testsuite/29_atomics/atomic_integral/operators/increment_neg.cc:
Likewise.
* testsuite/experimental/polymorphic_allocator/pmr_typedefs_string.cc:
Validate std::experimental::pmr::u8string.
* testsuite/experimental/string_view/typedefs.cc: Validate that the
u8string_view typedef is defined.
* testsuite/util/testsuite_common_types.h: Add char8_t, char16_t and
char32_t to the typelists.

From-SVN: r269089

Disambiguate __gnu_cxx::append_ partial specialization

This patch corrects ambiguous partial specializations of
typelist::detail::append_.  Previously, neither append_<chain<Hd, Tl>,
Typelist_Chain> nor append_<Typelist_Chain, null_type> was a better
match for append_<chain<Hd, Tl>, null_type>.

2019-02-22  Tom Honermann  <tom@honermann.net>

* include/ext/typelist.h: Constrain a partial specialization of
typelist::detail::append_ to only match chain<T1,T2>.

From-SVN: r269088

re PR go/89406 (Go testing leaves many temporary directories in /tmp around)

PR go/89406
    net: remove unixgram test sockets

    Backport https://golang.org/cl/163277 from the master library.

    Updates https://gcc.gnu.org/PR89406

    Reviewed-on: https://go-review.googlesource.com/c/163200

From-SVN: r269087

cmd/go: remove work directory on usage error

    Backport https://golang.org/cl/163237 from the master library:

        Ensure that cmd/go consistently calls base.Exit rather than os.Exit,
        so that we don't incorrectly leave the work directory around on exit.

        Test this by modifying the testsuite to run all the tests with TMPDIR
        set to a temporary directory, and then check that no files are left
        behind in that temporary directory. Adjust a couple of tests to make
        this approach work.

    Updates https://gcc.gnu.org/PR89406

    Reviewed-on: https://go-review.googlesource.com/c/163198

From-SVN: r269086

Daily bump.

From-SVN: r269085

extend.texi (__clear_cache): Correct signature.

gcc/ChangeLog:

* doc/extend.texi (__clear_cache): Correct signature.

libgcc/ChangeLog:

* libgcc2.h (__clear_cache): Correct signature.
* libgcc2.c (__clear_cache): Same.

gcc/testsuite/ChangeLog:

* gcc.dg/Wbuiltin-declaration-mismatch-12.c: New test.

From-SVN: r269082

PR c++/89422 - ICE with -g and lambda in default arg in template.

Here, we were trying to instantiate the default argument before setting
DECL_FRIEND_CONTEXT, so that the instantiated lambda ended up being treated
as part of the S template, which confused dwarf2out.

* pt.c (tsubst_function_decl): SET_DECL_FRIEND_CONTEXT sooner.

From-SVN: r269081

PR c++/88419 - C++17 ICE with class template arg deduction.

Just like in make_constrained_auto, we need to defer setting TYPE_CANONICAL
until we've set fields that will affect structural_comptypes.

* pt.c (make_template_placeholder): Set TYPE_CANONICAL after
CLASS_PLACEHOLDER_TEMPLATE.

From-SVN: r269080

re PR go/89170 (FAIL: net/http)

PR go/89170
* varasm.c (decode_addr_const): Call lookup_constant_def rather
than output_constant_def.
(add_constant_to_table): New static function.
(output_constant_def): Call add_constant_to_table.
(tree_output_constant_def): Likewise.

From-SVN: r269079

re PR c++/89285 (ICE after casting the this pointer in the constructor in C++17 mode)

PR c++/89285
* builtins.c (fold_builtin_arith_overflow): If first two args are
INTEGER_CSTs, set intres and ovfres to constants rather than calls
to ifn.

* constexpr.c (struct constexpr_fundef): Add parms and result members.
(retrieve_constexpr_fundef): Adjust for the above change.
(register_constexpr_fundef): Save constexpr body with copy_fn,
temporarily set DECL_CONTEXT on DECL_RESULT before that.
(get_fundef_copy): Change FUN argument to FUNDEF with
constexpr_fundef * type, grab body and parms/result out of
constexpr_fundef struct and temporarily change it for copy_fn calls
too.
(cxx_eval_builtin_function_call): For __builtin_FUNCTION temporarily
adjust current_function_decl from ctx->call context.  Test
!potential_constant_expression instead of !is_constant_expression.
(cxx_bind_parameters_in_call): Grab parameters from new_call.  Undo
convert_for_arg_passing changes for TREE_ADDRESSABLE type passing.
(cxx_eval_call_expression): Adjust get_fundef_copy caller.
(cxx_eval_conditional_expression): For IF_STMT, allow then or else
operands to be NULL.
(label_matches): Handle BREAK_STMT and CONTINUE_STMT.
(cxx_eval_loop_expr): Add support for FOR_STMT, WHILE_STMT and DO_STMT.
(cxx_eval_switch_expr): Add support for SWITCH_STMT.
(cxx_eval_constant_expression): Handle IF_STMT, FOR_STMT, WHILE_STMT,
DO_STMT, CONTINUE_STMT, SWITCH_STMT, BREAK_STMT and CONTINUE_STMT.
For SIZEOF_EXPR, recurse on the result of fold_sizeof_expr.  Ignore
DECL_EXPR with USING_DECL operand.
* lambda.c (maybe_add_lambda_conv_op): Build thisarg using
build_int_cst to make it a valid constant expression.

* g++.dg/ubsan/vptr-4.C: Expect reinterpret_cast errors.
* g++.dg/cpp1y/constexpr-84192.C (f2): Adjust expected diagnostics.
* g++.dg/cpp1y/constexpr-70265-2.C (foo): Adjust expected line of
diagnostics.
* g++.dg/cpp1y/constexpr-89285.C: New test.
* g++.dg/cpp0x/constexpr-arith-overflow.C (add, sub, mul): Ifdef out
for C++11.
(TEST_ADD, TEST_SUB, TEST_MUL): Define to Assert (true) for C++11.
* g++.dg/cpp0x/constexpr-arith-overflow2.C: New test.

From-SVN: r269078

PR libstdc++/89416 fix __is_move_insertable trait

The common base class for __is_move_insertable and __is_copy_insertable
instantiates both the copy and move tests, when only one is needed. The
unneeded one might cause errors outside the immediate context.

The solution used in this patch is to replace them with alias templates,
which will only be instantiated as needed.

PR libstdc++/89416
* include/bits/alloc_traits.h (__is_alloc_insertable_impl): Replace
class template with class. Replace move and copy member types with
member alias templates, so they are only instantiated when needed.
(__is_copy_insertable, __is_move_insertable): Adjust base class.
* testsuite/23_containers/vector/modifiers/push_back/89130.cc: Enable
test for C++11/14/17 as well.
* testsuite/23_containers/vector/modifiers/push_back/89416.cc: New
test.

From-SVN: r269075

i386: Replace -fcf-protection with -fcf-protection=branch

Since -mindirect-branch is incompatible with -fcf-protection=return and
-fcf-protection, replace -fcf-protection with -fcf-protection=branch.

PR target/87412
* gcc.target/i386/indirect-thunk-attr-14.c: Replace
-fcf-protection with -fcf-protection=branch.
* gcc.target/i386/indirect-thunk-attr-15.c: Likewise.
* gcc.target/i386/indirect-thunk-attr-16.c: Likewise.
* gcc.target/i386/indirect-thunk-extern-8.c: Likewise.
* gcc.target/i386/indirect-thunk-extern-9.c: Likewise.
* gcc.target/i386/indirect-thunk-extern-10.c: Likewise.

From-SVN: r269074

PR c++/88690 - C++17 ICE with empty base in aggregate.

Base fields for empty bases appear in initialization order, which may not be
the same as layout order.  If they also show up in a CONSTRUCTOR in that
order, output_constructor_regular_field aborts because it understandably
doesn't want to go backwards.  I also considered making o_c_r_f more
tolerant of the case where the out-of-order field has fieldsize 0, and so no
actual data needs to be emitted, but we might as well avoid adding an
element to the CONSTRUCTOR in the first place.

* typeck2.c (process_init_constructor_record): Skip trivial
initialization of an empty base.

From-SVN: r269073

Fix pr88850-2 test

Like the pr88850.c test, also fix the pr88850-2.c testcase which
was failing in hardfp environments.

Committed as obvious.

    gcc/testsuite/
* gcc.target/arm/pr88850-2.c: Block -mfloat-abi override.
* gcc.target/arm/pr88850.c: Use -mfloat-abi=softfp.

From-SVN: r269072

dump-parse-tree.c (debug): Implement for gfc_expr *, gfc_typespec *, gfc_typespec and gfc_symbol *.

2019-02-21  Thomas Koenig  <tkoenig@gcc.gnu.org>

    * dump-parse-tree.c (debug): Implement for gfc_expr *,
    gfc_typespec *, gfc_typespec and gfc_symbol *.

From-SVN: r269071

re PR fortran/86119 (Intrinsic len has wrong type if used within select type for a class(*) string)

2019-02-21  Thomas Koenig  <tkoenig@gcc.gnu.org>

    PR fortran/86119
    * class.c (gfc_get_len_component): Add argument k for kind.
    If the kind of the resulting expression is not equal to k,
    convert it.
    * gfortran.h (gfc_len_component): Adjust prototype.
    * simplify.c (gfc_simplify_len): Pass kind to
    gfc_get_len_component.

2019-02-21  Thomas Koenig  <tkoenig@gcc.gnu.org>

    PR fortran/86119
    * gfortran.dg/warn_conversion_11.f90: New test.

From-SVN: r269070

i386: Check -mindirect-branch/-mfunction-return with -fcf-protection

Issue an error when -mindirect-branch or -mfunction-return are used
with incompatible -fcf-protection.

gcc/

PR target/87412
* config/i386/i386.c (ix86_set_indirect_branch_type): Issue an
error for -mindirect-branch/-mfunction-return with incompatible
-fcf-protection.

gcc/testsuite/

PR target/87412
* gcc.target/i386/pr87412-1.c: New file.
* gcc.target/i386/pr87412-2.c: Likewise.
* gcc.target/i386/pr87412-3.c: Likewise.
* gcc.target/i386/pr87412-4.c: Likewise.

From-SVN: r269068

re PR bootstrap/88714 (bootstrap comparison failure on armv7l since r265398)

PR bootstrap/88714
* constraints.md (q): Remove.
* config/arm/ldrdstrd.md (*arm_ldrd, *arm_strd): Use rk constraint
instead of q.

From-SVN: r269067

[omp] Move NE_EXPR handling to omp_adjust_for_condition

2019-02-21  Martin Jambor  <mjambor@suse.cz>

PR hsa/89302
* omp-general.c (omp_extract_for_data): Removed a duplicate call
to omp_adjust_for_condition, moved NE_EXPR code_cond processing...
(omp_adjust_for_condition): ...here.  Added necessary parameters.
* omp-general.h (omp_adjust_for_condition): Updated declaration.
* omp-grid.c (grid_attempt_target_gridification): Adjust to pass
proper values to new parameters of omp_adjust_for_condition.

From-SVN: r269066

re PR c++/89392 (ICE in bitmap_bit_p, at bitmap.c:978)

2019-02-21  Richard Biener  <rguenther@suse.de>

PR middle-end/89392
cp/
* vtable-class-hierarchy.c (vtv_generate_init_routine): Do not
make symtab process new functions here.

From-SVN: r269065

PR c++/87921 - wrong error with inline static data member.

c_parse_final_cleanups checks DECL_IN_AGGR_P to avoid trying to emit a
static data member that has not been defined.  The inline variable patch
changed that to exempt inline variables.  But in this case we haven't
instantiated the variable yet, so we really don't have a definition.  This
patch changes inline variable handling such that DECL_IN_AGGR_P is not set
for a defined inline variable, so we can remove all the checks of
DECL_INLINE_VAR_P after DECL_IN_AGGR_P.

With that change we were failing on a static data member that had been
instantiated due to a use before we got around to processing it in
instantiate_class_template; we should detect that and avoid all the
finish_static_data_member_decl processing, which assumes that it is the
first time we're seeing the variable.

* decl2.c (finish_static_data_member_decl): Don't set DECL_IN_AGGR_P
for a non-template inline variable.  Do nothing for an
already-instantiated variable.
(c_parse_final_cleanups): Check DECL_IN_AGGR_P without
DECL_INLINE_VAR_P.
* decl.c (check_initializer): Likewise.
(make_rtl_for_nonlocal_decl): Likewise.
* pt.c (instantiate_decl): Likewise.
* typeck2.c (store_init_value): Likewise.

From-SVN: r269064

re PR go/89407 (go bootstrap failure on s390x starting with r268941)

PR go/89407
    internal/cpu: use #ifdef __s390x__ in C code

    Patch by Jakub Jelinek.

    Fixes https://gcc.gnu.org/PR89407

    Reviewed-on: https://go-review.googlesource.com/c/163297

From-SVN: r269063

Daily bump.

From-SVN: r269062

re PR c++/89403 (ICE in maybe_clone_body, at cp/optimize.c:693)

PR c++/89403
* decl2.c (c_parse_final_cleanups): Move TREE_ASM_WRITTEN setting
for flag_syntax_only from here...
* semantics.c (expand_or_defer_fn_1): ... here.

* g++.dg/cpp0x/pr89403.C: New test.

From-SVN: r269059

re PR c++/89405 (ICE in import_export_decl, at cp/decl2.c:2959)

PR c++/89405
* decl.c (maybe_commonize_var): When clearing TREE_PUBLIC and
DECL_COMMON, set DECL_INTERFACE_KNOWN.

* g++.dg/cpp1z/inline-var5.C: New test.

From-SVN: r269058

re PR middle-end/89412 (gcc ICE in simplify_subreg, at simplify-rtx.c:6273 on i686-linux-gnu)

PR middle-end/89412
* expr.c (expand_assignment): If result is a MEM, use change_address
instead of simplify_gen_subreg.

* gcc.c-torture/compile/pr89412.c: New test.

From-SVN: r269057

re PR middle-end/89091 (ICE: Segmentation fault (in tree_class_check))

PR middle-end/89091
* fold-const.c (decode_field_reference): Return NULL_TREE if
lang_hooks.types.type_for_size returns NULL.  Check it before
overwriting *exp_.  Use return NULL_TREE instead of return 0.

* gcc.dg/torture/pr89091.c: New test.

Co-Authored-By: David Malcolm <dmalcolm@redhat.com>
From-SVN: r269056

re PR tree-optimization/88074 (g++ hangs on math expression)

PR middle-end/88074
PR middle-end/89415
* toplev.c (do_compile): Double the emin/emax exponents to workaround
buggy mpc_norm.

* gcc.dg/pr88074-2.c: New test.

From-SVN: r269055

re PR target/89397 (ICE in build_call_expr_loc_array at gcc/tree.c:11563 since r229082)

PR target/89397
* config/i386/i386.c (ix86_atomic_assign_expand_fenv): Check
TARGET_SSE in addition to TARGET_SSE_MATH.

(ix86_excess_precision): Ditto.
(ix86_float_exceptions_rounding_supported_p): Ditto.
(use_rsqrt_p): Ditto.
* config/i386/sse.md (rsqrt<mode>2): Ditto.

From-SVN: r269054

linux-unwind.h (alpha_fallback_frame_state): Cast 'mcontext_t *' &rt_->uc.uc_mcontext to 'struct sigcontext *'.

* config/alpha/linux-unwind.h (alpha_fallback_frame_state):
Cast 'mcontext_t *' &rt_->uc.uc_mcontext to 'struct sigcontext *'.

From-SVN: r269053

re PR c++/89336 (internal compiler error when compiling a constexpr function)

PR c++/89336
* constexpr.c (cxx_eval_store_expression): Diagnose changing of active
union member for -std=c++17 and earlier.

* g++.dg/cpp1y/constexpr-89336-3.C: New test.

From-SVN: r269052

Fix ICE with #line directive (PR c/89410)

PR c/89410 reports various issues with #line directives with very
large numbers; one of them is an ICE inside diagnostic-show-locus.c
when emitting a diagnostic at line 0xffffffff.

The issue is that the arithmetic in layout::calculate_line_spans to
determine if two line spans are sufficiently close to consolidate
was using the unsigned 32-bit linenum_type, which was overflowing
when comparing the line for the expanded location with those of
the location range (all on line 0xffffffff), leading to it
erroneously adding two spans for the same line, leading to an
assertion failure.

This patch fixes the ICE by generalizing the use of long long in
line-map.h's comparison function for linenum_type into a new
linenum_arith_t typedef, and using it here.

Doing so uncovered a second problem: the loop to print the lines
within the line_span for this case is infinite: looping from
0xfffffff upwards, overflowing to 0, and then never becoming
greater than 0xfffffff.  The patch fixes this by using linenum_arith_t
there also.

gcc/ChangeLog:
PR c/89410
* diagnostic-show-locus.c (layout::calculate_line_spans): Use
linenum_arith_t when determining if two adjacent line spans are
close enough to merge.
(diagnostic_show_locus): Use linenum_arith_t when iterating over
lines within each line_span.

gcc/testsuite/ChangeLog:
PR c/89410
* gcc.dg/pr89410-1.c: New test.
* gcc.dg/pr89410-2.c: New test.

libcpp/ChangeLog:
PR c/89410
* include/line-map.h (linenum_arith_t): New typedef.
(compare): Use it.

From-SVN: r269050

compiler: fix a typo in comments

    Reviewed-on: https://go-review.googlesource.com/c/163097

From-SVN: r269049

PR c++/87513 - 'sorry' mangling PMF template-id.

Here build_offset_ref calls build_qualified_name to make a SCOPE_REF because
the dependent template arguments make type_dependent_expression_p (member)
true.  We could probably work hard to prevent this, but it doesn't seem
necessary, and it's easy to fix write_expression to handle the result.

* mangle.c (write_expression): Handle SCOPE_REF to BASELINK.

From-SVN: r269048

* g++.old-deja/g++.robertl/eb92.C: Handle warning as well.

From-SVN: r269047

PR c++/88380 - wrong-code with flexible array and NSDMI.

Here 'skipped' was set to -1 to force an explicit initializer for 'uninit'
before the initializer for 'initialized', and so we also tried to emit an
explicit initializer for the flexible array, for which build_zero_init
returns error_mark_node.  We should ignore flexarrays even when
skipped < 0.

* typeck2.c (process_init_constructor_record): Skip flexarrays.

From-SVN: r269046

PR c++/88572 - wrong handling of braces on scalar init.

* decl.c (reshape_init_r): Allow braces around scalar initializer
within aggregate init.  Reject double braced-init of scalar
variable.

From-SVN: r269045