[testsuite] Add missing fenv effective target to fp-int-convert-timode* tests.

2019-01-25  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/
* doc/sourcebuild.texi (Environment attributes): Add fenv and
fenv_exceptions description.

2019-01-25  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/testsuite/
* lib/target-supports.exp (check_effective_target_fenv): New.
* gcc.dg/torture/fp-int-convert-float128-timode-3.c: Add missing
fenv effective target.
* gcc.dg/torture/fp-int-convert-timode-1.c: Likewise.
* gcc.dg/torture/fp-int-convert-timode-2.c: Likewise.
* gcc.dg/torture/fp-int-convert-timode-3.c: Likewise.
* gcc.dg/torture/fp-int-convert-timode-4.c: Likewise.

From-SVN: r268266

[PATCH][AArch64] Fix generation of tst (PR87763)

The TST instruction no longer matches in all cases due to changes in
Combine.  The fix is simple, we now need to allow a subreg as well when
selecting the cc_mode.  This fixes the tst_5.c and tst_6.c failures.

AArch64 regress & bootstrap OK.

PR rtl-optimization/87763
* config/aarch64/aarch64.c (aarch64_select_cc_mode):
Allow SUBREG when matching CC_NZmode compare.

From-SVN: r268265

re PR tree-optimization/89049 (Unexpected vectorization)

2019-01-25  Richard Biener  <rguenther@suse.de>

PR tree-optimization/89049
* tree-vect-loop.c (vect_compute_single_scalar_iteration_cost):
Look at the pattern stmt to determine if the stmt is vectorized.

From-SVN: r268264

[AArch64][SVE] Handle register-register pred_movs

pred_mov<mode> is defined for predicated loads and stores, where
exactly one of the operands is a register.  However, the instruction
condition only checked for "one" rather than "exactly one", and
Prathamesh found a case in which combine could fold a predicated
pattern to an all-register pred_mov<mode>.  The constraints would
then force one of the registers to memory.

This patch splits all-register forms into a normal move as soon
as possible, but also adds an all-register alternative in case the
instruction doesn't get split before RA (or in case the RA can use
inheritance to avoid a reload).

The testcase for this will be added to aarch64/sve-acle-branch.

2018-01-25  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-sve.md (*pred_mov<mode>)
(pred_mov<mode>): Handle all-register forms using both a new
alternative and a split.

From-SVN: r268263

re PR tree-optimization/86865 (Wrong code w/ -O2 -floop-parallelize-all -fstack-reuse=none -fwrapv -fno-tree-ch -fno-tree-dce -fno-tree-dominator-opts -fno-tree-loop-ivcanon)

2019-01-25  Richard Biener  <rguenther@suse.de>

PR tree-optimization/86865
* graphite-scop-detection.c (scop_detection::can_represent_loop):
Reject non-do-while loops.

* gcc.dg/graphite/pr86865.c: New testcase.
* gcc.dg/graphite/pr69728.c: XFAIL.
* gcc.dg/graphite/scop-21.c: Likewise.

From-SVN: r268257

Daily bump.

From-SVN: r268256

altivec.md (build_vector_mask_for_load): Use MEM_P.

* config/rs6000/altivec.md (build_vector_mask_for_load): Use MEM_P.
* config/rs6000/constraints.md (Q constraint): Use REG_P.
* config/rs6000/darwin.h (PREFERRED_RELOAD_CLASS): Use SYMBOL_REF_P.
* config/rs6000/freebsd64.h (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P): Use
SYMBOL_REF_P, CONST_INT_P and CONST_DOUBLE_P.
* config/rs6000/linux64.h (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P): Likewise.
* config/rs6000/predicates.md (altivec_register_operand, vint_operand,
vsx_register_operand, vsx_reg_sfsubreg_ok, vfloat_operand,
vlogical_operand, gpc_reg_operand, int_reg_operand,
int_reg_operand_not_pseudo): Use SUBREG_P and HARD_REGISTER_P.
(ca_operand, base_reg_operand, htm_spr_reg_operand, cc_reg_operand,
cc_reg_not_cr0_operand, input_operand): Use SUBREG_P.
(save_world_operation, restore_world_operation, lmw_operation,
stmw_operation): Use MEM_P and REG_P.
(tie_operand): Use MEM_P.
(vrsave_operation, crsave_operation): Use REG_P.
(mfcr_operation, mtcrf_operation): Use REG_P and CONST_INT_P.
(fpr_reg_operand): Use SUBREG_P and HARD_REGISTER_NUM_P.
(quad_int_reg_operand): Use HARD_REGISTER_NUM_P.
(call_operand): Use HARD_REGISTER_P.
(indexed_or_indirect_operand, altivec_indexed_or_indirect_operand):
Use CONST_INT_P.
(lwa_operand): Use SUBREG_P, REG_P and CONST_INT_P.
* config/rs6000/rs6000-p8swap.c (insn_is_load_p, insn_is_store_p,
quad_aligned_load_p, replace_swapped_aligned_store,
recombine_lvx_pattern, replace_swapped_aligned_load,
recombine_stvx_pattern): Use MEM_P.
(const_load_sequence_p, adjust_vperm, replace_swapped_load_constant):
Use MEM_P and SYMBOL_REF_P.
(rtx_is_swappable_p): Use REG_P and CONST_INT_P.
(insn_is_swappable_p): Use REG_P and MEM_P.
(insn_is_swap_p, (alignment_mask): Use CONST_INT_P.
* config/rs6000/rs6000-string.c (expand_block_clear, expand_block_move):
Use CONST_INT_P.
* config/rs6000/rs6000.c (rs6000_secondary_reload, rs6000_emit_cmove):
Use CONST_DOUBLE_P.
(rs6000_output_move_128bit): Use CONST_DOUBLE_P, CONST_INT_P and
CONST_WIDE_INT_P.
(rs6000_legitimize_address): Use CONST_DOUBLE_P, CONST_INT_P,
CONST_WIDE_INT_P, REG_P and SYMBOL_REF_P.
(rs6000_emit_move): Use CONST_DOUBLE_P, CONST_INT_P, HARD_REGISTER_P,
HARD_REGISTER_NUM_P, MEM_P, REG_P, SUBREG_P, SYMBOL_REF_P and
reg_or_subregno:
(output_toc): Use CONST_DOUBLE_P, CONST_INT_P and SYMBOL_REF_P.
(easy_altivec_constant, rs6000_legitimate_offset_address_p,
rs6000_mode_dependent_address, rs6000_expand_mtfsf_builtin,
rs6000_expand_set_fpscr_rn_builtin, rs6000_expand_set_fpscr_drn_builtin,
rs6000_expand_unop_builtin, INT_P, rs6000_generate_compare,
rs6000_machopic_legitimize_pic_address, rs6000_split_logical_inner,
rs6000_split_logical_di): Use CONST_INT_P.
(rs6000_legitimize_reload_address): Use CONST_INT_P, HARD_REGISTER_P,
REG_P and SYMBOL_REF_P.
(setup_incoming_varargs, rs6000_rtx_costs): Use CONST_INT_P and MEM_P.
(print_operand): Use CONST_INT_P, MEM_P and REG_P.
(virtual_stack_registers_memory_p, rs6000_legitimate_address_p,
mems_ok_for_quad_peep): Use CONST_INT_P and REG_P.
(rs6000_secondary_reload_memory): Use CONST_INT_P and SUBREG_P.
(small_data_operand, print_operand_address): Use CONST_INT_P and
SYMBOL_REF_P.
(split_stack_arg_pointer_used_p): Use HARD_REGISTER_P.
(rs6000_init_hard_regno_mode_ok, direct_move_p):
Use HARD_REGISTER_NUM_P.
(rs6000_secondary_reload_gpr): Use HARD_REGISTER_NUM_P and MEM_P.
(rs6000_secondary_reload_class): Use HARD_REGISTER_NUM_P, REG_P,
SUBREG_P and SYMBOL_REF_P.
(register_to_reg_type, rs6000_secondary_reload_inner): Use SUBREG_P
and HARD_REGISTER_NUM_P.
(rs6000_adjust_vec_address): Use HARD_REGISTER_NUM_P and
reg_or_subregno.
(rs6000_adjust_cost, find_mem_ref): Use MEM_P.
(macho_lo_sum_memory_operand, rs6000_eliminate_indexed_memrefs): Use
MEM_P and REG_P.
(legitimate_indirect_address_p, legitimate_lo_sum_address_p,
registers_ok_for_quad_peep, rs6000_output_function_epilogue,
find_addr_reg): Use REG_P.
(altivec_expand_vec_perm_const): Use REG_P and SUBREG_P.
(rs6000_emit_le_vsx_move): Use SUBREG_P.
(offsettable_ok_by_alignment, constant_pool_expr_p,
legitimate_small_data_p, rs6000_output_dwarf_dtprel,
rs6000_delegitimize_address, rs6000_const_not_ok_for_debug_p,
rs6000_cannot_force_const_mem, rs6000_output_addr_const_extra,
rs6000_assemble_integer, create_TOC_reference,
rs6000_emit_allocate_stack, rs6000_xcoff_encode_section_info,
rs6000_call_aix, rs6000_call_aix): Use SYMBOL_REF_P.
(rs6000_split_vec_extract_var): Use reg_or_subregno.
* config/rs6000/rtems.h (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P): Use
CONST_DOUBLE_P, CONST_INT_P and SYMBOL_REF_P.
* config/rs6000/sysv4.h (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P): Likewise.
* config/rs6000/xcoff.h (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P): Likewise.
* config/rs6000/rs6000.h (RS6000_SYMBOL_REF_TLS_P): Use SYMBOL_REF_P.
(REGNO_OK_FOR_INDEX_P, REGNO_OK_FOR_BASE_P): Use HARD_REGISTER_NUM_P.
(INT_REG_OK_FOR_INDEX_P, INT_REG_OK_FOR_BASE_P): Use HARD_REGISTER_P.
(CONSTANT_ADDRESS_P): Use CONST_INT_P and SYMBOL_REF_P.
* config/rs6000/rs6000.md (define_expands strlensi, mod<mode>3
and cbranch<mode>4): Use CONST_INT_P.
(multiple define_splits): Use REG_P and SUBREG_P.
(define_expands call, call_value): Use MEM_P.
(define_expands sibcall, sibcall_value): Use CONST_INT_P and  MEM_P.
(define insn *mtcrfsi): Use CONST_INT_P and REG_P.
* config/rs6000/vsx.md (*vsx_le_perm_load_<mode>,
*vsx_le_perm_load_v8hi, *vsx_le_perm_load_v16qi): Use HARD_REGISTER_P
and HARD_REGISTER_NUM_P.
(multiple define_splits): Use HARD_REGISTER_NUM_P.

From-SVN: r268253

PR c++/89001 - mangling of reference temporaries

It used to be the case that the mangled name of a reference temporary didn't
need to be standardized, because all access would be through the reference.
But now constant expressions can look through references and so different
translation units need to agree on the address of a temporary in the
initializer of a reference with vague linkage.

* cp-tree.h (struct saved_scope): Add ref_temp_count.
(current_ref_temp_count): New macro.
* mangle.c (mangle_ref_init_variable): Use it.
* typeck2.c (store_init_value): Clear it.
* call.c (make_temporary_var_for_ref_to_temp): Copy public and
comdat.

From-SVN: r268252

PR c/86125 - missing -Wbuiltin-declaration-mismatch on a mismatched return type

PR c/86125 - missing -Wbuiltin-declaration-mismatch on a mismatched return type
PR middle-end/86308 - ICE in verify_gimple calling index() with an invalid declaration
PR c/86125 - missing -Wbuiltin-declaration-mismatch on a mismatched return type
PR c/88886 - [9 Regression] ice in get_constant, at c-family/c-format.c:292

gcc/c/ChangeLog:

PR c/86125
PR c/88886
PR middle-end/86308
* c-decl.c (match_builtin_function_types): Add arguments.
(diagnose_mismatched_decls): Diagnose mismatched declarations
of built-ins more strictly.

gcc/testsuite/ChangeLog:

PR c/86125
PR c/88886
PR middle-end/86308
* gcc.dg/Wbuiltin-declaration-mismatch-6.c: New test.
* gcc.dg/Wbuiltin-declaration-mismatch-7.c: New test.
* gcc.dg/Wbuiltin-declaration-mismatch-8.c: New test.
* gcc.dg/Wbuiltin-declaration-mismatch-9.c: New test.
* gcc.dg/Wbuiltin-declaration-mismatch-10.c: New test.
* gcc.dg/builtins-69.c: New test.
* gcc.dg/Wint-conversion-2.c: Add expected warning.
* gcc.c-torture/execute/eeprof-1.c: Adjust function signatures.

From-SVN: r268251

ChangeLog: Correct PR classification.

* ChangeLog: Correct PR classification.
* testsuite/ChangeLog: Ditto.

From-SVN: r268250

re PR rtl-optimization/88948 (ICE in elimination_costs_in_insn, at reload1.c:3640 since r264148)

PR target/88948
* rtl.h (prepare_copy_insn): New prototype.
* gcse.c (prepare_copy_insn): New function, split out from
process_insert_insn.
(process_insert_insn): Use prepare_copy_insn.
* store-motion.c (replace_store_insn): Use prepare_copy_insn
instead of gen_move_insn.

testsuite/ChangeLog:

PR target/88948
* gcc.target/i386/pr88948.c: New test.

From-SVN: r268249

re PR debug/89006 (New note: non-delegitimized UNSPEC UNSPEC_SET_GOT (14) found in variable location since r267638)

PR debug/89006
* config/i386/i386.c (ix86_pic_register_p): Return true for
UNSPEC_SET_GOT too.

* g++.dg/debug/pr89006.C: New test.

From-SVN: r268248

re PR tree-optimization/88964 (ICE in wide_int_to_tree_1, at tree.c:1561)

PR tree-optimization/88964
* gimple-loop-interchange.cc (loop_cand::analyze_induction_var): Also
punt if HONOR_SNANS (chrec).

From-SVN: r268247

re PR middle-end/89015 (ICE in lookup_decl_in_outer_ctx, at omp-low.c:3480)

PR middle-end/89015
* tree-nested.c (convert_nonlocal_reference_stmt,
convert_local_reference_stmt, convert_tramp_reference_stmt,
convert_gimple_call) <case GIMPLE_OMP_TEAMS>: Treat
gimple_omp_teams_host teams stmts like GIMPLE_OMP_PARALLEL
or GIMPLE_OMP_TASK.

* gcc.dg/gomp/pr89015.c: New test.

From-SVN: r268246

re PR c++/88976 (ICE in fold_convert_loc, at fold-const.c:2552)

PR c++/88976
* c-typeck.c (c_finish_omp_cancel): Diagnose more than one if
on #pragma omp cancel with different modifiers.

* semantics.c (finish_omp_cancel): Diagnose more than one if
on #pragma omp cancel with different modifiers.  Use
maybe_convert_cond when not in template or build_x_binary_op
otherwise.

* c-c++-common/gomp/cancel-2.c: New test.
* gcc.dg/gomp/cancel-1.c: New test.
* g++.dg/gomp/cancel-1.C: New test.
* g++.dg/gomp/cancel-2.C: New test.
* g++.dg/gomp/cancel-3.C: New test.

From-SVN: r268245

re PR tree-optimization/89027 (ICE: verify_gimple failed (Error: non-trivial conversion at assignment))

PR tree-optimization/89027
* tree-inline.c (add_clobbers_to_eh_landing_pad): Don't add clobbers
for "omp simd array" variables.

* gfortran.dg/gomp/pr89027.f90: New test.

From-SVN: r268243

Mitigation for PR target/88469 on arm-based systems bootstrapping with gcc-6/7/8

This patch, for gcc 8/9 is a mitigation patch for PR target/88469
where gcc-6/7/8 miscompile a structure whose alignment is dominated by
a 64-bit bitfield member.  Since the PCS rules for such a type must
ignore any overalignment of the base type we cannot address this by
simply adding a larger alignment to the class.  We can, however, force
the alignment of the bit-field itself and GCC will handle that as
desired.

PR target/88469
* profile-count.h (profile_count): On ARM systems using GCC 6/7/8
force the alignment of m_val.

From-SVN: r268240

PR libstdc++/88840 delay evaluation of constant until type is complete

Clang fails to compile std::vector<Incomplete> because the static member
__use_relocate cannot be evaluated for an incomplete type. Replace with
a static member function that will not be odr-used until needed, by
which point the type must be complete.

PR libstdc++/88840
* include/bits/stl_vector.h (vector::__use_relocate): Replace static
data member with static member function _S_use_relocate().
* include/bits/vector.tcc (vector::reserve, vector::_M_realloc_insert)
(vector::_M_default_append): Use _S_use_relocate() instead of
__use_relocate.

From-SVN: r268239

Fix failing test due to inconsistent strcmp results

* testsuite/27_io/filesystem/path/compare/strings.cc: Only compare
sign of results.

From-SVN: r268238

[nvptx, libgomp] Fix memleak in GOMP_OFFLOAD_fini_device

I wrote a test-case:
...
int
main (void)
{
  for (unsigned i = 0; i < 128; ++i)
    {
      acc_init (acc_device_nvidia);
      acc_shutdown (acc_device_nvidia);
    }

  return 0;
}
...
and ran it under valgrind.  The only leak location reported with a frequency
of 128, was the allocation of ptx_devices in nvptx_init.

Fix this by freeing ptx_devices in GOMP_OFFLOAD_fini_device, once
instantiated_devices drops to 0.

2019-01-24  Tom de Vries  <tdevries@suse.de>

* plugin/plugin-nvptx.c (GOMP_OFFLOAD_fini_device): Free ptx_devices
once instantiated_devices drops to 0.

From-SVN: r268237

re PR lto/87187 (FAIL: gfortran.dg/short_circuiting_3.f90   -g -flto  (internal compiler error) on darwin)

2019-01-24  Richard Biener  <rguenther@suse.de>

PR lto/87187
* tree-streamer-out.c (write_ts_decl_common_tree_pointers):
When in "legacy" debug mode make sure to reset self-origins.

From-SVN: r268236

Fix broken filename for .gcda files starting with '..' (PR gcov-profile/88994).

2019-01-24  Martin Liska  <mliska@suse.cz>

PR gcov-profile/88994
* gcov-io.c (mangle_path): Do not allocate a bigger buffer,
result will be always smaller or equal to the original.
* gcov.c (mangle_name): Fix else branch where we should
also copy to PTR and shift the pointer.

From-SVN: r268233

re PR fortran/88929 (ICE on building MPICH 3.2 with GCC 9 with ISO_Fortran_binding)

2019-01-24  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/88929
* trans-array.c (gfc_conv_descriptor_elem_len): New function.
* trans-array.h : Add prototype for above.
* trans-expr.c (gfc_conv_gfc_desc_to_cfi_desc): Take account of
assumed rank arrays being flagged by rank = -1 in expressions.
Intent in arrays need a pointer to a copy of the data to be
assigned to the descriptor passed for conversion. This should
then be freed, together with the CFI descriptor on return from
the C call.

2019-01-24  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/88929
* gfortran.dg/ISO_Fortran_binding_3.f90 : New test
* gfortran.dg/ISO_Fortran_binding_3.c : Subsidiary source.

From-SVN: r268232

re PR fortran/88929 (ICE on building MPICH 3.2 with GCC 9 with ISO_Fortran_binding)

2019-01-24  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/88929
* trans-array.c (gfc_conv_descriptor_elem_len): New function.
* trans-array.h : Add prototype for above.
* trans-expr.c (gfc_conv_gfc_desc_to_cfi_desc): Take account of
assumed rank arrays being flagged by rank = -1 in expressions.
Intent in arrays need a pointer to a copy of the data to be
assigned to the descriptor passed for conversion. This should
then be freed, together with the CFI descriptor on return from
the C call.

2019-01-24  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/88929
* gfortran.dg/ISO_Fortran_binding_3.f90 : New test
* gfortran.dg/ISO_Fortran_binding_3.c : Subsidiary source.

From-SVN: r268231

runtime: install SIGURG handler on C created threads

    Precise stack scan uses SIGURG to trigger a stack scan. We need
    to have Go signal handler installed for SIGURG.

    Reviewed-on: https://go-review.googlesource.com/c/159097

From-SVN: r268230

fix comments typo.

gcc/ChangeLog

2019-01-24  Xiong Hu Luo  <luoxhu@linux.vnet.ibm.com>

* tree-ssa-dom.c (test_for_singularity): fix a comment typo.
* vr-values.c (find_case_label_ranges): fix a comment typo.

From-SVN: r268229

fix tab alignment issue.

ChangeLog

2019-01-24  Xiong Hu Luo  <luoxhu@linux.vnet.ibm.com>

* ChangeLog: replace space with tab.
* MAINTAINERS: delete 1 tab to keep alignment.

From-SVN: r268228

Daily bump.

From-SVN: r268225

re PR target/85711 (ICE in aarch64_classify_address, at config/aarch64/aarch64.c:5678)

2019-01-23  Bin Cheng  <bin.cheng@arm.com>
    Steve Ellcey <sellcey@marvell.com>

PR target/85711
* recog.c (address_operand): Return false on wrong mode for address.
(constrain_operands): Check for mode with 'p' constraint.

Co-Authored-By: Steve Ellcey <sellcey@marvell.com>
From-SVN: r268219

libgcc2.c: Correct DI/TI -> SF/DF conversions

FSTYPE FUNC (DWtype u) in libgcc2.c, which converts DI/TI to SF/DF, has

  /* No leading bits means u == minimum.  */
  if (count == 0)
    return -(Wtype_MAXp1_F * (Wtype_MAXp1_F / 2));

in the third case (where actually count == 0 only means the high part is
minimum).  It should be:

  /* No leading bits means u == minimum.  */
  if (count == 0)
    return Wtype_MAXp1_F * (FSTYPE) (hi | ((UWtype) u != 0));

instead.

gcc/testsuite/

2019-01-23  H.J. Lu  <hongjiu.lu@intel.com>

PR libgcc/88931
* gcc.dg/torture/fp-int-convert-timode-1.c: New test.
* gcc.dg/torture/fp-int-convert-timode-2.c: Likewise.
* gcc.dg/torture/fp-int-convert-timode-3.c: Likewise.
* gcc.dg/torture/fp-int-convert-timode-4.c: Likewise.

libgcc/

2019-01-23  Joseph Myers  <joseph@codesourcery.com>

PR libgcc/88931
* libgcc2.c (FSTYPE FUNC (DWtype u)): Correct no leading bits
case.

From-SVN: r268216

* g++.target/i386/pr88998.C (main): Use FP constant in the assert.

From-SVN: r268198

re PR target/88998 (bad codegen with mmx instructions for unordered_map)

PR target/88998
* config/i386/sse.md (sse2_cvtpi2pd): Add SSE alternatives.
Disparage MMX alternative.
(sse2_cvtpd2pi): Ditto.
(sse2_cvttpd2pi): Ditto.

testsuite/ChangeLog:

PR target/88998
* g++.target/i386/pr88998.c: New test.

From-SVN: r268195

PR c++/88757 - qualified name treated wrongly as type.

* parser.c (cp_parser_direct_declarator): don't treat qualified-ids
in parameter-list as types if name lookup for declarator-id didn't
find one or more function templates.

* g++.dg/cpp0x/dependent2.c: new test.
* g++.dg/cpp2a/typename10.c: remove dg-error.
* g++.dg/cpp2a/typename12.c: new test.
* g++.dg/template/static30.c: remove dg-error.

From-SVN: r268192

aarch64: fix use-after-free in -march=native (PR driver/89014)

Running:
  $ valgrind ./xgcc -B. -c test.c -march=native
on aarch64 shows a use-after-free in host_detect_local_cpu due
to the std::string result of aarch64_get_extension_string_for_isa_flags
only living until immediately after a c_str call.

This leads to corrupt "-march=" values being passed to cc1.

This patch fixes the use-after-free, though it appears to also need
Tamar's patch here:
  https://gcc.gnu.org/ml/gcc-patches/2018-12/msg01302.html
in order to generate valid values for cc1.  This may have worked by
accident in the past, if the corrupt "-march=" value happened to be
0-terminated in the "right" place; with this patch it now appears
to reliably break without Tamar's patch.

gcc/ChangeLog:
PR driver/89014
* config/aarch64/driver-aarch64.c (host_detect_local_cpu): Fix
use-after-free of the result of
aarch64_get_extension_string_for_isa_flags.

From-SVN: r268189

re PR c/44715 (Break in increment expression of "for" statement inconsistent with g++)

PR c/44715
* cp-gimplify.c (genericize_cp_loop): Call begin_bc_block only
after genericizing cond and incr expressions.

* doc/extend.texi: Document break and continue behavior in
statement expressions.

* c-c++-common/pr44715.c: New test.

From-SVN: r268188

re PR c++/88984 (ICE in genericize_switch_stmt, at cp/cp-gimplify.c:377)

PR c++/88984
* cp-gimplify.c (genericize_switch_stmt): Move cond genericization
before the begin_bc_block call.

* c-c++-common/pr88984.c: New test.

From-SVN: r268187

re PR middle-end/89008 (O2 and O1 results differ for simple test)

2019-01-23  Richard Biener  <rguenther@suse.de>

PR tree-optimization/89008
* tree-ssa-reassoc.c (eliminate_using_constants): For * 0 do
not leave another stray operand.

* gcc.dg/torture/pr89008.c: New testcase.

From-SVN: r268186

PR c++/87893 - constexpr ctor ICE on ARM.

PR c++/88293 - ICE with comma expression.
* constexpr.c (initialized_type): Don't shortcut non-void type.
Handle COMPOUND_EXPR.
(cxx_eval_outermost_constant_expr): Return early for void type.

From-SVN: r268185

* BASE-VER: Bump to 9.0.1.

From-SVN: r268184

re PR sanitizer/89010 (Update URLs in libsanitizer/README.gcc)

PR sanitizer/89010
* libsanitizer/README.gcc: Update to current https URLs.

From-SVN: r268183

cgraphunit.c (cgraph_node::expand_thunk): When expanding a GIMPLE thunk that returns by reference...

* cgraphunit.c (cgraph_node::expand_thunk): When expanding a GIMPLE
thunk that returns by reference, use the type of the return object
of the thunk instead of that of the alias to build the dereference.

From-SVN: r268182

[ARC] atomics: Add operand to DMB instruction

Atomics use DMB instruction to enforce ordering of loads/stores.
Currently gcc generates DMB w/o any arg which is a no-op. Fix that by
generating DMB 3 which enforces R+W ordering. It is stricter than what
acq/rel expect, but there's no other way.

gcc/

2019-xx-xx  Vineet Gupta <vgupta@synopsys.com>

       * config/arc/atomic.md: Add operand to DMB instruction

From-SVN: r268181

[libbacktrace] Use size_t for low_offset/high_offset fields of struct unit

2019-01-23  Tom de Vries  <tdevries@suse.de>

* dwarf.c (struct unit): Use size_t for low_offset/high_offset fields.
(units_search, find_unit): Use size_t for offset.
(build_address_map): Use size_t for unit_offset.

From-SVN: r268180

re PR tree-optimization/88964 (ICE in wide_int_to_tree_1, at tree.c:1561)

PR tree-optimization/88964
* gimple-loop-interchange.cc (loop_cand::analyze_induction_var): Use
build_zero_cst instead of build_int_cst.  Return false for loop
invariants which honor signed zeros.

* gfortran.dg/pr88964.f90: New test.

From-SVN: r268179

[nvptx, libgomp] Fix cuMemAlloc with size zero

Consider test-case:
...
int
main (void)
{
  #pragma acc parallel async
  ;
  #pragma acc parallel async
  ;
  #pragma acc wait

  return 0;
}
...

This fails with:
...
libgomp: cuMemAlloc error: invalid argument
Segmentation fault (core dumped)
...
The cuMemAlloc error is due to the fact that we're try to allocate 0 bytes.

Fix this by preventing calling map_push with size zero argument in nvptx_exec.

This also has the consequence that for the abort-1.c test-case, we end up
calling cuMemFree during map_fini for the struct cuda_map allocated in
map_init, which fails because an abort happened.  Fix this by calling
cuMemFree with CUDA_CALL_NOCHECK in cuda_map_destroy.

2019-01-23  Tom de Vries  <tdevries@suse.de>

PR target/PR88946
* plugin/plugin-nvptx.c (cuda_map_destroy): Use CUDA_CALL_NOCHECK for
cuMemFree.
(nvptx_exec): Don't call map_push if mapnum == 0.
* testsuite/libgomp.oacc-c-c++-common/pr88946.c: New test.

From-SVN: r268178

[nvptx, libgomp] Fix assert (!s->map->active) in map_fini

There are currently two situations where this assert triggers:
...
libgomp/plugin/plugin-nvptx.c: map_fini: Assertion `!s->map->active' failed.
...

First, in abort-1.c, a parallel region triggering an abort:
...
int
main (void)
{
  #pragma acc parallel
  abort ();

  return 0;
}
...

The abort is detected in nvptx_exec as the CUDA_ERROR_ILLEGAL_INSTRUCTION
return status of the cuStreamSynchronize call after kernel launch, which is
then handled by calling non-returning function GOMP_PLUGIN_fatal.
Consequently, the map_pop in nvptx_exec that in case of cuStreamSynchronize
success would remove or inactive the element added by the map_push earlier in
nvptx_exec, does not trigger.  With the element no longer active, but still
marked active and a member of s->map,  we run into the assert during
GOMP_OFFLOAD_fini_device, which is triggered from atexit handler
gomp_target_fini (which is triggered by the GOMP_PLUGIN_fatal mentioned above
calling exit).

Second, in pr88941.c, an async parallel region without wait:
...
int
main (void)
{
  #pragma acc parallel async
  ;

  /* no #pragma acc wait */
  return 0;
}
...

Because nvptx_exec is handling an async region, it does not call map_pop for
the element added by map_push, but schedules an kernel execution completion
event to call map_pop.  Again, we run into the assert during
GOMP_OFFLOAD_fini_device, which is triggered from atexit handler
gomp_target_fini, but the exit in this case is triggered by returning from main.
So either the kernel is still running, or the kernel has completed but the
corresponding event that is supposed to call map_pop is stuck in the event
queue, waiting for an event_gc.

Fix this by removing the assert, and skipping the freeing of device memory if
the map is still marked active (though in the async case, this is more a
workaround than an fix).

2019-01-23  Tom de Vries  <tdevries@suse.de>

PR target/88941
PR target/88939
* plugin/plugin-nvptx.c (cuda_map_destroy): Handle map->active case.
(map_fini): Remove "assert (!s->map->active)".
* testsuite/libgomp.oacc-c-c++-common/pr88941.c: New test.

From-SVN: r268177

[nvptx, libgomp] Fix map_push

The map field of a struct ptx_stream is a FIFO.  The FIFO is implemented as a
single linked list, with pop-from-the-front semantics.

The function map_pop pops an element, either by:
- deallocating the element, if there is more than one element
- or marking the element inactive, if there's only one element

The responsibility of map_push is to push an element to the back, as well as
selecting the element to push, by:
- allocating an element, or
- reusing the element at the front if inactive and big enough, or
- dropping the element at the front if inactive and not big enough, and
  allocating one that's big enough

The current implemention gets at least the first and most basic scenario wrong:

> map = cuda_map_create (size);

We create an element, and assign it to map.

> for (t = s->map; t->next != NULL; t = t->next)
>   ;

We determine the last element in the fifo.

> t->next = map;

We append the new element.

> s->map = map;

But here, we throw away the rest of the FIFO, and declare the FIFO to be just
the new element.

This problem causes the test-case asyncwait-1.c to fail intermittently on some
systems.  The pr87835.c test-case added here is a a minimized and modified
version of asyncwait-1.c (avoiding the kernel construct) that is more likely to
fail.

Fix this by rewriting map_pop more robustly, by:
- seperating the function in two phases: select element, push element
- when reusing or dropping an element, making sure that the element is cleanly
  popped from the queue
- rewriting the push element part in such a way that it can handle all cases
  without needing if statements, such that each line is exercised for each of
  the three cases.

2019-01-23  Tom de Vries  <tdevries@suse.de>

PR target/87835
* plugin/plugin-nvptx.c (map_push): Fix adding of allocated element.
* testsuite/libgomp.oacc-c-c++-common/pr87835.c: New test.

From-SVN: r268176

Warray-bounds-2.c: Include headers only if they exist.

gcc/testsuite/ChangeLog:

* c-c++-common/Warray-bounds-2.c: Include headers only if they exist.
* c-c++-common/Warray-bounds-3.c: Make xfails conditional on target
non_strict_align.
* c-c++-common/Wrestrict-2.c: Include headers only if they exist.
* c-c++-common/Wrestrict.c: Make xfails conditional on target
non_strict_align.

From-SVN: r268175

Daily bump.

From-SVN: r268174

Fix -fsplit-paths documentation

It is enabled at -O3, but the doc currently says -O2.  This fixes it.

* doc/invoke.texi (-fsplit-paths): This is enabled by default at -O3.

From-SVN: r268170

c-c++-common/gomp/pr60823-1.c: Change aarch64-*-* target to aarch64*-*-* target.

2018-01-22  Steve Ellcey  <sellcey@marvell.com>

c-c++-common/gomp/pr60823-1.c: Change aarch64-*-* target
to aarch64*-*-* target.
c-c++-common/gomp/pr60823-3.c: Ditto.
g++.dg/gomp/declare-simd-1.C: Ditto.
g++.dg/gomp/declare-simd-3.C: Ditto.
g++.dg/gomp/declare-simd-4.C: Ditto.
g++.dg/gomp/declare-simd-7.C: Ditto.
g++.dg/gomp/pr88182.C: Ditto.
gcc.dg/gomp/declare-simd-1.c: Ditto.
gcc.dg/gomp/declare-simd-3.c: Ditto.
gcc.dg/gomp/pr59669-2.c: Ditto.
gcc.dg/gomp/pr87895-1.c: Ditto.
gcc.dg/gomp/simd-clones-2.c: Ditto.
gfortran.dg/gomp/declare-simd-2.f90: Ditto.
gfortran.dg/gomp/pr79154-1.f90: Ditto.
gfortran.dg/gomp/pr83977.f90: Ditto.

From-SVN: r268168

d/dmd: Merge dmd upstream e21c07e84

Fixes bootstrap regression introduced by the previous merge.

Reviewed-on: https://github.com/dlang/dmd/pull/9283

From-SVN: r268167

re PR target/88965 (powerpc64le vector builtin hits ICE in verify_gimple)

PR target/88965
* config/rs6000/rs6000.c: Include tree-vrp.h and tree-ssanames.h.
(rs6000_gimple_fold_builtin): If MEM_REF address doesn't satisfy
is_gimple_mem_ref_addr predicate, force it into a SSA_NAME first.

* gcc.target/powerpc/pr88965.c: New test.

From-SVN: r268166

re PR middle-end/88968 (Stack overflow in gimplify_expr)

PR middle-end/88968
* gimplify.c (gimplify_omp_atomic): Handle bitfield atomics with
non-integral DECL_BIT_FIELD_REPRESENTATIVEs.

* c-omp.c (c_finish_omp_atomic): For bitfield atomics, update type
variable after using BIT_FIELD_REF.

* c-c++-common/gomp/atomic-23.c: New test.

From-SVN: r268165

re PR target/87064 (libgomp.oacc-fortran/reduction-3.f90 fails starting with r263751)

PR target/87064
* config/rs6000/vsx.md (*vsx_reduc_<VEC_reduc_name>_v2df_scalar):
Disable for little endian.

From-SVN: r268164

re PR fortran/88579 (Calculating power of powers of two)

2019-01-22  Harald Anlauf  <anlauf@gmx.de>

PR fortran/88579
* trans-expr.c (gfc_conv_power_op): Handle cases of (2**e) ** integer
and (- 2**e) ** integer.

2019-01-22  Harald Anlauf  <anlauf@gmx.de>

PR fortran/88579
* gfortran.dg/power_8.f90: New test.

From-SVN: r268163

pr87906_0.C: Add dg-require-effective-target fpic.

2019-01-22  Sandra Loosemore  <sandra@codesourcery.com>

gcc/testsuite/
* g++.dg/lto/pr87906_0.C: Add dg-require-effective-target fpic.
* g++.dg/vec-init-1.C: Likewise.
* gcc.dg/pr87793.c: Likewise.

From-SVN: r268162

pr86397-1.C: Add -fdelete-null-pointer-checks.

2019-01-22  Sandra Loosemore  <sandra@codesourcery.com>

gcc/testsuite/
* g++.dg/cpp0x/pr86397-1.C: Add -fdelete-null-pointer-checks.
* g++.dg/cpp0x/pr86397-2.C: Likewise.

From-SVN: r268161

[arm] Further fixes for PR88469

A bitfield that is exactly the same size as an integral type and
naturally aligned will have DECL_BIT_FIELD cleared.  So we need to
check DECL_BIT_FIELD_TYPE to be sure whether or not the underlying
type was declared with a bitfield declaration.

I've also added a test for bitfields that are based on overaligned types.

PR target/88469
gcc:
* config/arm/arm.c (arm_needs_double_word_align): Check
DECL_BIT_FIELD_TYPE.

gcc/testsuite:
* gcc.target/arm/aapcs/bitfield2.c: New test.
* gcc.target/arm/aapcs/bitfield3.c: New test.

From-SVN: r268160

Fix vect-nop-move.c test

Fix a failing test - changes in Combine mean the test now fails
eventhough the generated code is the same.  Given there are several
AArch64-specific tests for vec-select, remove the scanning of Combine
output.  Committed as trivial fix.

    testsuite/
PR rtl-optimization/87763
* gcc.dg/vect/vect-nop-move.c: Fix testcase on AArch64.

From-SVN: r268159

i386: Add mask2 to builtin_description

There are

struct builtin_description
{
  const HOST_WIDE_INT mask;
  const enum insn_code icode;
  const char *const name;
  const enum ix86_builtins code;
  const enum rtx_code comparison;
  const int flag;
};

Since "mask" is used for both ix86_isa_flags and ix86_isa_flags2, buitins
with both flags can't be handled easily.  This patch adds mask2 to
builtin_description to handle it properly.

2019-01-22  Hongtao Liu  <hongtao.liu@intel.com>
    H.J. Lu  <hongjiu.lu@intel.com>

PR target/88909
* config/i386/i386-builtin.def: Add mask2 to all builtin
initializations.  Merge ARGS2 and SPECIAL_ARGS2 into ARGS and
SPECIAL_ARGS.
* config/i386/i386.c (BDESC): Add mask2 to the definition.
(BDESC_FIRST): Likewise.
(define_builtin): Add an argument for mask2.  Updated to handle
both ix86_isa_flags and ix86_isa_flags2.
(define_builtin_const): Likewise.
(define_builtin_pure): Likewise.
(define_builtin2): Deleted.
(define_builtin_const2): Likewise.
(builtin_description): Add a member, mask2.
(bdesc_*): Add mask2 to builtin initializations.
(ix86_init_mmx_sse_builtins): Update calls to def_builtin,
def_builtin_const and def_builtin_pure.  Remove SPECIAL_ARGS2
support.
(ix86_get_builtin_func_type): Remove SPECIAL_ARGS2 support.

Co-Authored-By: H.J. Lu <hongjiu.lu@intel.com>
From-SVN: r268155

PR libstdc++/88740 Print assertion messages to stderr

PR libstdc++/88740
* testsuite/util/testsuite_hooks.h [stderr] (VERIFY): Use fprintf to
write to stderr instead of using printf.

From-SVN: r268154

golang.org/x/tools/go/analysis: pass "gccgo" to types.SizesFor

    For the gofrontend copy, change calls to types.SizesFor to pass
    "gccgo" rather than "gc".  Leave the asmdecl pass unchanged since that
    pass is gc-specific anyhow.

    This has been fixed in a better way in the external repo by
    https://golang.org/cl/158317 and friends, but that is not in 1.12, so
    use this approach for now.

    Reviewed-on: https://go-review.googlesource.com/c/158842

From-SVN: r268153

i386: Load external function address via GOT slot

With noplt attribute, we load the external function address via the GOT
slot so that linker won't create an PLT entry for extern function address.

gcc/

PR target/88954
* config/i386/i386.c (ix86_force_load_from_GOT_p): Also check
noplt attribute.

gcc/testsuite/

PR target/88954
* gcc.target/i386/pr88954-1.c: New test.
* gcc.target/i386/pr88954-2.c: Likewise.

From-SVN: r268152

[arm] PR target/88469 fix incorrect argument passing with 64-bit bitfields

Unfortunately another PCS bug has come to light with the layout of
structs whose alignment is dominated by a 64-bit bitfield element.
Such fields in the type list appear to have alignment 1, but in
reality, for the purposes of alignment of the underlying structure,
the alignment is derived from the underlying bitfield's type.  We've
been getting this wrong since support for over-aligned record types
was added several releases back.  Worse still, the existing code may
generate unaligned memory accesses that may fault on some versions of
the architecture.

I've taken the opportunity to add a few more tests that check the
passing arguments with overalignment in the PCS.  Looking through the
existing tests it looked like they were really only checking
self-consistency and not the precise location of the arguments.

PR target/88469

gcc:
* config/arm/arm.c (arm_needs_doubleword_align): Return 2 if a record's
alignment is dominated by a bitfield with 64-bit aligned base type.
(arm_function_arg): Emit a warning if the alignment has changed since
earlier GCC releases.
(arm_function_arg_boundary): Likewise.
(arm_setup_incoming_varargs): Likewise.

gcc/testsuite:
* gcc.target/arm/aapcs/bitfield1.c: New test.
* gcc.target/arm/aapcs/overalign_rec1.c: New test.
* gcc.target/arm/aapcs/overalign_rec2.c: New test.
* gcc.target/arm/aapcs/overalign_rec3.c: New test.

From-SVN: r268151

array_function_5.f90: Fix a dg directive.

2019-01-22  Manfred Schwarb  <manfred99@gmx.ch>

* gfortran.dg/array_function_5.f90: Fix a dg directive.
* gfortran.dg/block_16.f08: Likewise.
* gfortran.dg/dec_structure_14.f90: Likewise.
* gfortran.dg/namelist_96.f90: Likewise.
* gfortran.dg/newunit_5.f90.f90: Moved to
* gfortran.dg/newunit_5.f90: here.
* gfortran.dg/pdt_28.f03: Likewise.
* gfortran.dg/spread_simplify_1.f90: Likewise.

From-SVN: r268148

re PR tree-optimization/88862 (ICE in extract_affine, at graphite-sese-to-poly.c:313)

2019-01-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/88862
* graphite-scop-detection.c
(scop_detection::graphite_can_represent_scev): Reject ADDR_EXPR.

From-SVN: r268147

Document AMD GCN.

2019-01-22  Andrew Stubbs  <ams@codesourcery.com>

* doc/extend.tex (AMD GCN Function Attributes): New section.
* doc/install.texi (amdgcn-unknown-amdhsa): New instructions.
* doc/invoke.texi (AMD GCN Options): New section.
* doc/md.texi (Constraints for Particular Machines): Add AMD GCN.

From-SVN: r268146

sparc.c (parc_delegitimize_address): Recognize the GOT register and decoded HIGH/LO_SUM combinations for labels in...

* config/sparc/sparc.c (parc_delegitimize_address): Recognize the GOT
register and decoded HIGH/LO_SUM combinations for labels in PIC mode.

From-SVN: r268145

Fix overwide shifts in PR/88903 testcase

gcc/testsuite/ChangeLog:

PR/tree-optimization 88903
* gcc.dg/vect/pr88903-1.c: Add explicit &.

From-SVN: r268144

re PR tree-optimization/88044 (gfortran.dg/transfer_intrinsic_3.f90 hangs after r266171)

PR tree-optimization/88044
* tree-ssa-loop-niter.c (number_of_iterations_cond): If condition
is false in the first iteration, but !every_iteration, return false
instead of true with niter->niter zero.

From-SVN: r268143

re PR lto/88422 (collect2.exe: fatal error: lto-wrapper returned 1 exit status: file not recognized: file truncated)

2019-01-22  Nidal Faour  <nidal.faour@wdc.com>

PR lto/88422
* simple-object.c (O_BINARY): Define if not already defined.
(simple_object_copy_lto_debug_sections): Create file in binary
mode.

From-SVN: r268141

re PR rtl-optimization/88904 (Basic block incorrectly skipped in jump threading.)

PR rtl-optimization/88904
* cfgcleanup.c (thread_jump): Verify cond2 doesn't mention
any nonequal registers before processing BB_END (b).

* gcc.c-torture/execute/pr88904.c: New test.

From-SVN: r268140

re PR target/88905 (ICE: in decompose, at rtl.h:2253 with -mabm and __builtin_popcountll)

PR target/88905
* optabs.c (add_equal_note): Add op0_mode argument, use it instead of
GET_MODE (op0).
(expand_binop_directly, expand_doubleword_clz,
expand_doubleword_popcount, expand_ctz, expand_ffs,
expand_unop_direct, maybe_emit_unop_insn): Adjust callers.

* gcc.dg/pr88905.c: New test.

From-SVN: r268139

re PR rtl-optimization/49429 (dse.c change (r175063) causes execution failures)

PR rtl-optimization/49429
PR target/49454
PR rtl-optimization/86334
PR target/88906
* expr.c (emit_block_move_hints): Move marking of MEM_EXPRs
addressable from here...
(emit_block_op_via_libcall): ... to here.

* gcc.target/i386/pr86334.c: New test.
* gcc.target/i386/pr88906.c: New test.

From-SVN: r268138

tree-vect-loop.c (vect_analyze_loop_operations): Use auto_vec for cost vector to fix memleak.

2019-01-22  Richard Biener  <rguenther@suse.de>

* tree-vect-loop.c (vect_analyze_loop_operations): Use
auto_vec for cost vector to fix memleak.
(vectorize_fold_left_reduction): Properly gather SLP defs.
(vectorizable_comparison): Do not swap operands to properly
gather SLP defs.

From-SVN: r268137

MAINTAINERS: add myself.

ChangLog:

2019-01-21  Jiufu Guo  <guojiufu@linux.ibm.com>

        * MAINTAINERS (Write After Approval): Add myself.

From-SVN: r268136

[RS6000] PR88614, output_operand: invalid %z value

The direct cause of this PR is the fact that tls_gdld_nomark didn't
handle indirect calls.  Also, most indirect calls were being optimised
back to direct calls anyway, due to tls_gdld_nomark not checking any
of the parallel elements except the first (plus the extra element that
distinguishes this call from normal calls).  There were other unwanted
substitutions too.

So this patch attacks the problem of handling special calls in a
different way.  Rather than adding another element to the call insn
parallel to distinguish -mno-tls-markers __tls_get_addr calls from any
other calls, we now inspect the second CALL arg.  Each
call_value_nonlocal and call_value_indirect insn now checks for the
tlsgd/ld unspecs when !TARGET_TLS_MARKERS and emits the arg setup
insns.  I disallow the local call patterns since we'll only see local
calls to __tls_get_addr in testcases, and it doesn't seem a good idea
to complicate the patterns just for a minor optimisation.  Sibling
call insns aren't used for libcalls, so none of these insns need to
change.

The patch also fixes a minor problem with -mno-tls-markers
__tls_get_addr calls causing a "li 3,0" instruction to be emitted
prior to the arg setup instructions, due to using a libcall with one
arg.  That isn't correct when the call insn itself sets up its arg.
Also, I've tidied the V4 secure-plt calls, generating them in
rs6000_call_sysv rather than by splitting in rs6000.md.  The
CALL_INSN_FUNCTION_USAGE added in edit_tls_call_insn is no longer
needed (since git commit 0a4b5c66df9).

On the subject of unwanted substitutions, I also saw a
_GLOBAL_OFFSET_TABLE_ symbol_ref being substituted for the GOT reg,
resulting in code like "addi 3,_GLOBAL_OFFSET_TABLE_,tls_ld@got@tlsld".
Fixed by the unspec_tls change.

PR target/88614
* config/rs6000/predicates.md (unspec_tls): Ensure GOT reg
stays a reg.  Allow a const_int.
* config/rs6000/rs6000-protos.h (rs6000_output_tlsargs): Declare.
* config/rs6000/rs6000.h (IS_V4_FP_ARGS): Define.
(IS_NOMARK_TLSGETADDR): Define.
* config/rs6000/rs6000.c (edit_tls_call_insn): Delete.
(rs6000_output_tlsargs): New function.
(rs6000_legitimize_tls_address): Don't say a !TARGET_TLS_MARKERS
__tls_get_addr call takes an arg.
(rs6000_call_sysv): Generate sysv4 secure plt call pattern here..
* config/rs6000/rs6000.md (call_nonlocal_sysv): ..rather than here,
delete split..
(call_value_nonlocal_sysv): ..or here, delete split.
(tls_gdld_nomark): Delete.
(call_value_indirect_nonlocal_sysv): Use unspec_tls as operand2
predicate.  Call rs6000_output_tlsargs.  Adjust length to suit.
(call_value_nonlocal_sysv): Likewise.
(call_value_nonlocal_sysv_secure): Likewise.
(call_value_nonlocal_aix): Likewise.
(call_value_indirect_aix): Likewise.
(call_value_indirect_elfv2): Likewise.
(call_value_local32, call_value_local64): Disable for no-mark tls.
(call_value_local_aix): Likewise.

From-SVN: r268135

Daily bump.

From-SVN: r268134

re PR go/88927 (Bootstrap failure on arm in libgo starting with r268084)

PR go/88927
    runtime, internal/cpu: fix build for ARM GNU/Linux

    Was failing with

    ../../../libgo/go/internal/cpu/cpu.go:138:2: error: reference to undefined name 'doinit'
      138 |  doinit()
          |  ^

    Fix it by adding in Go 1.12 internal/cpu/cpu_arm.go, and the code in
    runtime that initializes the values.

    Fixes https://gcc.gnu.org/PR88927.

    Reviewed-on: https://go-review.googlesource.com/c/158717

From-SVN: r268131

libgo: fix building, and some testing, on Solaris

    Restore some of the fixes that were applied to golang_org/x/net/lif
    but were lost when 1.12 moved the directory to internal/x/net/lif.

    Add support for reading /proc to fetch argc/argv/env for c-archive mode.

    Reviewed-on: https://go-review.googlesource.com/c/158640

From-SVN: r268130

compiler: stop iterator on error

    If there is an error reading or parsing an archive header, the
    Archive_iterator code would return a dummy header but would not mark
    itself as done.  The effect is that an invalid archive leads to an
    endless loop reading and re-reading the same archive header.  Avoid
    that by setting the offset to the end of the file, which will cause
    the iterator to == archive_end.

    No test since it doesn't seem worth constructing an invalid archive.

    Reviewed-on: https://go-review.googlesource.com/c/158217

From-SVN: r268129

re PR c++/88949 (ICE in expand_expr_real_1, at expr.c:10001 with -fopenmp)

PR c++/88949
* optimize.c (cxx_copy_decl): New function.
(clone_body): Use it instead of copy_decl_no_change.

* g++.dg/gomp/pr88949.C: New test.

From-SVN: r268127

2019-01-21  Dominique d'Humieres  <dominiq@gcc.gnu.org>

* Fix a typo in the previous commit.

From-SVN: r268126

class_66.f90: Fix a dg directive.

2019-01-21  Manfred Schwarb  <manfred99@gmx.ch>

* class_66.f90: Fix a dg directive.
* debug/pr35154-stabs.f: Likewise.
* dec_d_lines_3.f: Likewise.
* dec_d_lines_3.f: Likewise.
* dec_structure_12.f90: Likewise.
* dec_structure_15.f90: Likewise.
* deferred_character_31.f90: Likewise.
* dtio_31.f03: Likewise.
* dtio_32.f03: Likewise.
* extends_11.f03: Likewise.
* integer_plus.f90: Likewise.
* pdt_25.f03: Likewise.
* pr58968.f: Likewise.
* pr78259.f90: Likewise.
* vect/vect-2.f90: Likewise.
* matmul_const.f90: Likewise.

From-SVN: r268125

Merge dmd upstream 180465274

Reduces the memory footprint of the CTFE interpreter by replacing new
with emplacement new in many places.

gcc/d/ChangeLog:

2019-01-21  Iain Buclaw  <ibuclaw@gdcproject.org>

* d-frontend.cc (Compiler::paintAsType): Update for new signature.

From-SVN: r268124

re PR target/88938 (ICE in extract_insn, at recog.c:2304)

PR target/88938
* config/i386/i386.c (ix86_expand_builtin) [case IX86_BUILTIN_BEXTRI32,
case IX86_BUILTIN_BEXTRI64]: Sanitize operands.

testsuite/ChangeLog:

PR target/88938
* gcc.target/i386/pr88938.c: New test.

From-SVN: r268123

re PR sanitizer/88901 (ICE when using -fsanitize=pointer-compare)

PR sanitizer/88901
* typeck.c (cp_build_binary_op): Don't instrument
SANITIZE_POINTER_COMPARE if processing_template_decl.
(pointer_diff): Similarly for SANITIZE_POINTER_SUBTRACT.

* g++.dg/asan/pr88901.C: New test.

From-SVN: r268122

hash-map-tests.c (test_map_of_strings_to_int): Show how to use string contents as hash_map keys.

* hash-map-tests.c (test_map_of_strings_to_int): Show how to use
string contents as hash_map keys.

From-SVN: r268121

AArch64: Fix big-endian and ILP32 fail for simd-clone test.

gcc/testsuite/ChangeLog:

* g++.dg/vect/simd-clone-7.cc: Fix assembler scan.

--This line, a d those below, will be ignored--

M    gcc/testsuite/ChangeLog
M    gcc/testsuite/g++.dg/vect/simd-clone-7.cc

From-SVN: r268120

Makefile.am (noinst_LTLIBRARIES): Rename libzgcj_convience.la to libz_convenience.la.

2019-01-21  Iain Buclaw  <ibuclaw@gdcproject.org>

* Makefile.am (noinst_LTLIBRARIES): Rename libzgcj_convience.la to
libz_convenience.la.
* Makefile.in: Regenerate.
* configure.ac: Remove target_all.
* configure: Regenerate.

From-SVN: r268119

re PR c/88928 (ICE segfault in check_address_or_pointer_of_packed_member since r268075)

2019-01-21  Bernd Edlinger  <bernd.edlinger@hotmail.de>

        PR c/88928
        * c-warn.c (check_alignment_of_packed_member): Add a boolean parameter
        for rvalue context.  Handle rvalues correctly.  Use min_align_of_type
        instead of TYPE_ALIGN.
        (check_address_or_pointer_of_packed_member): Handle rvalues coorrectly.
        Use min_align_of_type instead of TYPE_ALIGN_UNIT.  Check for NULL
        pointer from TYPE_STUB_DECL.

testsuite:
2019-01-21  Bernd Edlinger  <bernd.edlinger@hotmail.de>

        PR c/88928
        * c-c++-common/Waddress-of-packed-member-1.c: New test case.
        * gcc.dg/pr88928.c: New test case.

From-SVN: r268118

utf-array.c: Allow wchar_t to be printed as {long ,short ,}{unsigned ,}int.

* gcc.dg/utf-array.c: Allow wchar_t to be printed as
{long ,short ,}{unsigned ,}int.

From-SVN: r268117

Fix leak in splay-tree

Philippe Waroquiers noticed a memory leak in gdb, which he tracked
down to a bug in splay-tree.  splay_tree_remove does not call the
`delete_key' function when it removes the old node; but it should.

I looked at every splay tree in GCC and there is only one that passes
a non-NULL delete function -- the one in lto.c.  That file does not
call splay_tree_remove.  So, I think this is safe to check in.

I re-ran the LTO tests to double check.

libiberty/
* splay-tree.c (splay_tree_remove): Delete the key if necessary.

From-SVN: r268116

re PR tree-optimization/88934 (ICE: verify_gimple failed (Error: mismatching comparison operand types))

2019-01-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/88934
* tree-vect-slp.c (vect_mask_constant_operand_p): Always look
at the possibly non-constant operand.
(vect_get_constant_vectors): Adjust.

* gfortran.dg/pr88934.f90: New testcase.

From-SVN: r268115

i386: Move Intel intrinsics head files to <immintrin.h>

According to Intel Intrinsics Guide:

https://software.intel.com/sites/landingpage/IntrinsicsGuide/

Intel intrinsics should be available by including <immintrin.h>.  This
patch moves remaining Intel intrinsics head files from <x86intrin.h> to
<immintrin.h>.

PR target/71659
* config/i386/adxintrin.h: Just check _IMMINTRIN_H_INCLUDED.
* config/i386/clflushoptintrin.h: Check _IMMINTRIN_H_INCLUDED
instead of _X86INTRIN_H_INCLUDED.
* onfig/i386/clwbintrin.h: Likewise.
* config/i386/pkuintrin.h: Likewise.
* config/i386/prfchwintrin.h: Likewise.
* config/i386/rdseedintrin.h: Likewise.
* config/i386/wbnoinvdintrin.h: Likewise.
* config/i386/xsavecintrin.h: Likewise.
* config/i386/xsavesintrin.h: Likewise.
* config/i386/fxsrintrin.h: Enable _IMMINTRIN_H_INCLUDED check.
* config/i386/xsaveintrin.h: Likewise.
* config/i386/xsaveoptintrin.h: Likewise.
* config/i386/x86intrin.h: Move "#include" <rdseedintrin.h>,
<prfchwintrin.h>, <fxsrintrin.h>, <xsaveintrin.h>,
<xsaveoptintrin.h>, <adxintrin.h>, <clwbintrin.h>,
<clflushoptintrin.h>, <xsavesintrin.h>, <xsavecintrin.h>,
<wbnoinvdintrin.h> and <pkuintrin.h> to ...
* config/i386/immintrin.h: Here.

From-SVN: r268113

re PR ipa/86590 (Codegen is poor when passing std::string by value with _GLIBCXX_EXTERN_TEMPLATE undefined)

PR libstdc++/86590
* include/bits/char_traits.h (__constant_string_p,
__constant_char_array_p): Use __builtin_is_constant_evaluated if
available.

From-SVN: r268112

Fix after P0600.

gcc/testsuite/ChangeLog
2019-02-20  Ulrich Drepper  <drepper@redhat.com>

Fix after P0600.
* g++.dg/init/new39.C: Don't just ignore result of new.

libstdc++/ChangeLog
2019-02-20  Ulrich Drepper  <drepper@redhat.com>

Implement C++20 P0600r1.
* include/backward/hash_map: Add nodiscard attribute to empty.
* include/backward/hash_set: Likewise.
* backward/hashtable.h: Likewise.
* include/bits/basic_string.h: Likewise.
* include/bits/forward_list.h: Likewise.
* include/bits/hashtable.h: Likewise.
* include/bits/regex.h: Likewise.
* include/bits/stl_deque.h: Likewise.
* include/bits/stl_list.h: Likewise.
* include/bits/stl_map.h: Likewise.
* include/bits/stl_multimap.h: Likewise.
* include/bits/stl_multiset.h: Likewise.
* include/bits/stl_queue.h: Likewise.
* include/bits/stl_set.h: Likewise.
* include/bits/stl_stack.h: Likewise.
* include/bits/stl_tree.h: Likewise.
* include/bits/stl_vector.h: Likewise.
* include/bits/unordered_map.h: Likewise.
* include/bits/unordered_set.h: Likewise.
* include/debug/array: Likewise.
* include/experimental/any: Likewise.
* include/experimental/bits/fs_path.h: Likewise.
* include/experimental/internet: Likewise.
* include/experimental/string_view: Likewise.
* include/ext/pb_ds/detail/bin_search_tree_/info_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/binary_heap_/binary_heap_.hpp:
Likewise.
* include/ext/pb_ds/detail/binary_heap_/info_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/cc_hash_table_map_/cc_ht_map_.hpp:
Likewise.
* include/ext/pb_ds/detail/cc_hash_table_map_/info_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/cc_hash_table_map_/size_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/gp_hash_table_map_/gp_ht_map_.hpp:
Likewise.
* include/ext/pb_ds/detail/gp_hash_table_map_/info_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/left_child_next_sibling_heap_/info_fn_imps.hpp:
Likewise.
*
include/ext/pb_ds/detail/left_child_next_sibling_heap_/left_child_next_sibling_heap_.hpp:
Likewise.
* include/ext/pb_ds/detail/list_update_map_/info_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/list_update_map_/lu_map_.hpp:
Likewise.
* include/ext/pb_ds/detail/ov_tree_map_/info_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/ov_tree_map_/ov_tree_map_.hp:
Likewise.
* include/ext/pb_ds/detail/pat_trie_/info_fn_imps.hpp:
Likewise.
* include/ext/pb_ds/detail/pat_trie_/pat_trie_.hpp:
Likewise.
* include/ext/pb_ds/detail/rc_binomial_heap_/rc.hpp:
Likewise.
* include/ext/pb_ds/detail/tree_trace_base.hpp: Likewise.
* include/ext/pb_ds/trie_policy.hpp: Likewise.
* include/ext/rope: Likewise.
* include/ext/slist: Likewise.
* include/ext/vstring.h: Likewise.
* include/profile/array: Likewise.
* include/std/array: Likewise.
* include/tr1/array: Likewise.
* include/tr1/hashtable.h: Likewise.
* include/tr1/regex: Likewise.
* include/tr2/dynamic_bitset: Likewise.
* include/bits/alloc_traits.h: Add nodiscard attribute to
allocate.
* include/experimental/memory_resource: Likewise.
* include/ext/alloc_traits.h: Likewise.
* include/ext/array_allocator.h: Likewise.
* include/ext/bitmap_allocator.h: Likewise.
* include/ext/debug_allocator.h: Likewise.
* include/ext/extptr_allocator.h: Likewise.
* include/ext/mt_allocator.h: Likewise.
* include/ext/new_allocator.h: Likewise.
* include/ext/pool_allocator.h: Likewise.
* include/ext/throw_allocator.h: Likewise.
* include/std/scoped_allocator: Likewise.
* libsupc++/eh_alloc.cc: Likewise.
* include/std/future: Add nodiscard attribute to async.
* libsupc++/new: Add nodiscard attribute to new.

From-SVN: r268111

Daily bump.

From-SVN: r268110

Limit AA walking in IPA summary generation

2019-01-20  Martin Jambor  <mjambor@suse.cz>

PR ipa/87615
* ipa-prop.h (struct ipa_func_body_info): Replaced field aa_walked
with aa_walk_budget.
* cgraph.h (ipa_polymorphic_call_context::get_dynamic_type): Add
aa_walk_budget_p parameter.
* ipa-fnsummary.c (unmodified_parm_1): New parameter fbi.  Limit AA
walk.  Updated all callers.
(unmodified_parm): New parameter fbi, pass it to unmodified_parm_1.
(eliminated_by_inlining_prob): New parameter fbi, pass it on to
unmodified_parm.
(will_be_nonconstant_expr_predicate): New parameter fbi, removed
parameter info.  Extract info from fbi.  Pass fbi to recursive calls
and to unmodified_parm.
(phi_result_unknown_predicate): New parameter fbi, removed parameter
info, updated call to will_be_nonconstant_expr_predicate.
(param_change_prob): New parameter fbi, limit AA walking.
(analyze_function_body): Initialize aa_walk_budget in fbi.  Update
calls to various above functions.
* ipa-polymorphic-call.c (get_dynamic_type): Add aa_walk_budget_p
parameter.  Use it to limit AA walking.
* ipa-prop.c (detect_type_change_from_memory_writes): New parameter
fbi, limit AA walk.
(detect_type_change): New parameter fbi, pass it on to
detect_type_change_from_memory_writes.
(detect_type_change_ssa): Likewise.
(aa_overwalked): Removed.
(parm_preserved_before_stmt_p): Assume fbi is never NULL, stream line
accordingly, adjust to the neew AA limiting scheme.
(parm_ref_data_preserved_p): Likewise.
(ipa_compute_jump_functions_for_edge): Adjust call to
get_dynamic_type.
(ipa_analyze_call_uses): Likewise.
(ipa_analyze_virtual_call_uses): Pass fbi to detect_type_change_ssa.
(ipa_analyze_node): Initialize aa_walk_budget.
(ipcp_transform_function): Likewise.
* tree-ssa-sccvn.c (eliminate_dom_walker::eliminate_stmt): Update call
to get_dynamic_type.

From-SVN: r268107