[Ada] Use declared type for deciding on SPARK pointer rules

2019-10-10  Piotr Trojanek  <trojanek@adacore.com>

gcc/ada/

* sem_prag.adb (Analyze_Global_In_Decl_Part): Simplify previous
test, just like in a recent commit we simplified a similar test
for Depends contract.

From-SVN: r276810

re PR target/88630 (Incorrect float negating together with convertion to int on ST-40)

gcc/
PR target/88630
* config/sh/sh.h (TARGET_FPU_SH4_300): New macro.
* config/sh/sh.c (sh_option_override): Enable fsca and fsrra insns
also for TARGET_FPU_SH4_300.
(sh_emit_mode_set): Check for TARGET_FPU_SH4_300 instead of
TARGET_SH4_300.
* config/sh/sh.md (toggle_pr): Add TARGET_FPU_SH4_300 condition.
(negsf2): Expand to either negsf2_fpscr or negsf2_no_fpscr.
(*negsf2_i): Split into ...
(negsf2_fpscr, negsf2_no_fpscr): ... these new patterns.
(abssf2): Expand to either abssf2_fpsc or abssf2_no_fpsc.
(**abssf2_i): Split into ...
(abssf2_fpscr, abssf2_no_fpscr): ... these new patterns.
(negdf2): Expand to either negdf2_fpscr or negdf2_no_fpscr.
(*negdf2_i): Split into ...
(negdf2_fpscr, negdf2_no_fpscr): ... these new patterns.
(absdf2): Expand to either absdf2_fpscr or absdf2_no_fpsc.
(**abssf2_i): Split into ...
(absdf2_fpscr, absdf2_no_fpscr): ... these new patterns.

From-SVN: r276809

Make comp_ptr_ttypes_real return bool.

comp_ptr_ttypes_real could use bool instead of int since it only returns
0 or 1.

* typeck.c (comp_ptr_ttypes_real): Change the return type to bool.
Use false instead of 0.

From-SVN: r276808

re PR middle-end/92046 (Command line options (that are per-functions) are affecting --params which are global.)

2019-10-10  Richard Biener  <rguenther@suse.de>

PR middle-end/92046
* opts.c (finish_options): Do not influence global --params
from options that are adjustable per function.
* tree-vect-data-refs.c (vect_enhance_data_refs_alignment):
Apply --param adjustment based on active cost-model.
* tree-ssa-phiopt.c (cond_if_else_store_replacement): Disable
further store-sinking when vectorization or if-conversion
are not enabled.

From-SVN: r276807

re PR middle-end/92037 (LTO bootstrap broken in selftests)

PR middle-end/92037
* cgraph.c (symbol_table_test::symbol_table_test): Use ggc_alloc
rather than ggc_alloc_cleared to alloc symbol table.
* toplev.c (general_init): Likewise.
* cgraph.h (symbol_table): Explicitly construct every field.

From-SVN: r276804

lto-common.c (unify_scc): Do not merge anonymous NAMESPACE_DECLs.

2019-10-10  Richard Biener  <rguenther@suse.de>

* lto-common.c (unify_scc): Do not merge anonymous NAMESPACE_DECLs.

From-SVN: r276796

S/390: Add support for z15 as CPU name.

So far z15 was identified as arch13. After the machine has been
announced we can now add the real name.

gcc/ChangeLog:

2019-10-10  Andreas Krebbel  <krebbel@linux.ibm.com>

* common/config/s390/s390-common.c (PF_ARCH13): Rename to...
(PF_Z15): ... this.
* config.gcc: Add z15 as option for --with-arch and --with-tune
configure switches.
* config/s390/s390-c.c (s390_resolve_overloaded_builtin): Add
error reporting for unsupported builtins.
* config/s390/s390-opts.h (enum processor_type): Rename
PROCESSOR_8561_ARCH13 to PROCESSOR_8561_Z15.
* config/s390/8561.md: Rename arch13 to z15 throughout the file.
* config/s390/driver-native.c (s390_host_detect_local_cpu):
Likewise.
* config/s390/s390-builtins.def: Likewise.
* config/s390/s390.c (processor_table): Add z15 as option and keep arch13 as alternative.
(s390_expand_builtin): Add missing check for unsupported builtins.
(s390_canonicalize_comparison): Rename TARGET_ARCH13 to TARGET_Z15.
(s390_rtx_costs): Likewise.
(s390_get_sched_attrmask): Rename arch13 to z15.
(s390_get_unit_mask): Likewise.
(s390_is_fpd): Likewise.
(s390_is_fxd): Likewise.
* config/s390/s390.h (enum processor_flags): Likewise.
* config/s390/s390.md: Likewise.
* config/s390/vector.md: Likewise.
* config/s390/vx-builtins.md: Likewise.
* config/s390/s390.opt: Add z15 to processor_type value.

From-SVN: r276792

[Fortran, OpenMP] Actually pass use_device_addr on to the middle end

        * trans-openmp.c (gfc_trans_omp_clauses): Actually pass
        use_device_addr on to the middle end.

From-SVN: r276791

S/390: PR91035 Fix call to __morestack

For the call to __morestack we use a special ABI in the S/390 back-end
which requires us to emit a parameter block to the .rodata section.
It contains the label whereto __morestack needs to return.  The
parameter block needs to be explicit in RTL since we also need to take
the address of it loaded into r1 in order to pass its address to
__morestack.  In order to express correctly what __morestack does its
RTX also contained the return label. Hence we had the return label to
occur twice in the insn stream.  This is problematic when it comes to
redirecting edges.  The correlation between these two occurrences of
the label cannot be expressed so when doing a redirect only the label
in the jump RTX gets modified while the parameter block label stays as
is.

The patch avoids having two instancs of the label by merging the
parameter block generation and the __morestack call RTX into one. By
doing this I could also get rid of the unspec which was required for
the parameter block generation so far.

gcc/ChangeLog:

2019-10-10  Andreas Krebbel  <krebbel@linux.ibm.com>

PR target/91035
* config/s390/s390-protos.h (s390_output_split_stack_data): Add
prototype.
* config/s390/s390.md (UNSPECV_SPLIT_STACK_DATA): Remove.
("split_stack_data", "split_stack_call")
("split_stack_call_<mode>", "split_stack_cond_call")
("split_stack_cond_call_<mode>"): Remove.
("@split_stack_call<mode>", "@split_stack_cond_call<mode>"): New
insn definition.
* config/s390/s390.c (s390_output_split_stack_data): New function.
(s390_expand_split_stack_prologue): Use the merged expander.

From-SVN: r276790

c-common.h (c_omp_check_context_selector, [...]): Declare.

c-family/
* c-common.h (c_omp_check_context_selector,
c_omp_get_context_selector): Declare.
* c-omp.c (c_omp_declare_simd_clauses_to_numbers): Fix spelling
in diagnostic message.
(c_omp_check_context_selector, c_omp_get_context_selector): New
functions.
* c-attribs.c (c_common_attribute_table): Add "omp declare variant"
attribute.
(handle_omp_declare_variant_attribute): New function.
c/
* c-parser.c (c_parser_omp_all_clauses): Add NESTED_P argument, if
true, terminate processing on closing paren and don't skip to end of
pragma line.
(c_parser_omp_declare_simd): Handle also declare variant.
(omp_construct_selectors, omp_device_selectors,
omp_implementation_selectors, omp_user_selectors): New variables.
(c_parser_omp_context_selector,
c_parser_omp_context_selector_specification,
c_finish_omp_declare_variant): New functions.
(c_finish_omp_declare_simd): Handle both declare simd and
declare variant.
(c_parser_omp_declare): Handle declare variant.
cp/
* parser.h (struct cp_omp_declare_simd_data): Add variant_p member.
* parser.c (cp_ensure_no_omp_declare_simd): Handle both declare simd
and declare variant.
(cp_parser_oacc_all_clauses): Formatting fix.
(cp_parser_omp_all_clauses): Add NESTED_P argument, if true, terminate
processing on closing paren and don't skip to end of pragma line.
(cp_parser_omp_declare_simd): Add VARIANT_P argument.  Handle also
declare variant.
(omp_construct_selectors, omp_device_selectors,
omp_implementation_selectors, omp_user_selectors): New variables.
(cp_parser_omp_context_selector,
cp_parser_omp_context_selector_specification,
cp_finish_omp_declare_variant): New functions.
(cp_parser_late_parsing_omp_declare_simd): Handle also declare variant.
(cp_parser_omp_declare): Handle declare variant.
testsuite/
* c-c++-common/gomp/declare-variant-1.c: New test.
* c-c++-common/gomp/declare-variant-2.c: New test.
* c-c++-common/gomp/declare-variant-3.c: New test.
* g++.dg/gomp/this-1.C: Adjust for diagnostic message spelling fix.
* gcc.dg/gomp/declare-variant-1.c: New test.
* gcc.dg/gomp/declare-variant-2.c: New test.

From-SVN: r276789

* cp-tree.h (template_info_decl_check): Check ENABLE_TREE_CHECKING.

From-SVN: r276787

Daily bump.

From-SVN: r276786

PR tree-optimization/90879 - fold zero-equality of strcmp between a longer string and a smaller array

gcc/c-family/ChangeLog:

PR tree-optimization/90879
* c.opt (-Wstring-compare): New option.

gcc/testsuite/ChangeLog:

PR tree-optimization/90879
* gcc.dg/Wstring-compare-2.c: New test.
* gcc.dg/Wstring-compare.c: New test.
* gcc.dg/strcmpopt_3.c: Scan the optmized dump instead of strlen.
* gcc.dg/strcmpopt_6.c: New test.
* gcc.dg/strlenopt-65.c: Remove uinnecessary declarations, add
test cases.
* gcc.dg/strlenopt-66.c: Run it.
* gcc.dg/strlenopt-68.c: New test.

gcc/ChangeLog:

PR tree-optimization/90879
* builtins.c (check_access): Avoid using maxbound when null.
* calls.c (maybe_warn_nonstring_arg): Adjust to get_range_strlen change.
* doc/invoke.texi (-Wstring-compare): Document new warning option.
* gimple-fold.c (get_range_strlen_tree): Make setting maxbound
conditional.
(get_range_strlen): Overwrite initial maxbound when non-null.
* gimple-ssa-sprintf.c (get_string_length): Adjust to get_range_strlen
changes.
* tree-ssa-strlen.c (maybe_diag_stxncpy_trunc): Same.
(used_only_for_zero_equality): New function.
(handle_builtin_memcmp): Call it.
(determine_min_objsize): Return an integer instead of tree.
(get_len_or_size, strxcmp_eqz_result): New functions.
(maybe_warn_pointless_strcmp): New function.
(handle_builtin_string_cmp): Call it.  Fold zero-equality of strcmp
between a longer string and a smaller array.
(get_range_strlen_dynamic): Overwrite initial maxbound when non-null.

From-SVN: r276773

Implement C++20 P0388R4, DR 1307, and DR 330.

This patch implements P0388R4, Permit conversions to arrays of unknown bound,
<http://wg21.link/p0388r4>.  CWG 393 allowed references to arrays of unknown
bound and this C++20 feature allows conversions like

  void f(int(&)[]);
  int arr[1];

  void g() { f(arr); }
  int(&r)[] = arr;

The proposal seemed fairly straightforward but it turned out to be quite
shifty.  I found out that I needed to implement DR 2352 (done), and also
DR 1307 (done in this patch).  The latter DR added wording for
list-initialization ranking of references to arrays which this proposal
extends.  DR 330 was also implemented in this patch.

PR c++/91364 - P0388R4: Permit conversions to arrays of unknown bound.
PR c++/69531 - DR 1307: Differently bounded array parameters.
PR c++/88128 - DR 330: Qual convs and pointers to arrays of pointers.
* call.c (build_array_conv): Build ck_identity at the beginning
of the conversion.
(standard_conversion): Pass bounds_none to comp_ptr_ttypes_const.
(maybe_warn_array_conv): New.
(convert_like_real): Call it.  Add an error message about converting
from arrays of unknown bounds.
(conv_get_original_expr): New.
(nelts_initialized_by_list_init): New.
(conv_binds_to_array_of_unknown_bound): New.
(compare_ics): Implement list-initialization ranking based on
array sizes, as specified in DR 1307 and P0388R.
* cp-tree.h (comp_ptr_ttypes_const): Adjust declaration.
(compare_bounds_t): New enum.
* typeck.c (comp_array_types): New bool and compare_bounds_t
parameters.  Use them.
(structural_comptypes): Adjust the call to comp_array_types.
(similar_type_p): Handle ARRAY_TYPE.
(build_const_cast_1): Pass bounds_none to comp_ptr_ttypes_const.
(comp_ptr_ttypes_real): Don't check cv-quals of ARRAY_TYPEs.  Use
comp_array_types to compare array types.  Look through arrays as per
DR 330.
(comp_ptr_ttypes_const): Use comp_array_types to compare array types.
Look through arrays as per DR 330.

* g++.dg/conversion/qual1.C: New test.
* g++.dg/conversion/qual2.C: New test.
* g++.dg/conversion/qual3.C: New test.
* g++.dg/conversion/ref2.C: New test.
* g++.dg/conversion/ref3.C: New test.
* g++.dg/cpp0x/initlist-array3.C: Remove dg-error.
* g++.dg/cpp0x/initlist-array7.C: New test.
* g++.dg/cpp0x/initlist-array8.C: New test.
* g++.dg/cpp2a/array-conv1.C: New test.
* g++.dg/cpp2a/array-conv10.C: New test.
* g++.dg/cpp2a/array-conv11.C: New test.
* g++.dg/cpp2a/array-conv12.C: New test.
* g++.dg/cpp2a/array-conv13.C: New test.
* g++.dg/cpp2a/array-conv14.C: New test.
* g++.dg/cpp2a/array-conv15.C: New test.
* g++.dg/cpp2a/array-conv16.C: New test.
* g++.dg/cpp2a/array-conv17.C: New test.
* g++.dg/cpp2a/array-conv2.C: New test.
* g++.dg/cpp2a/array-conv3.C: New test.
* g++.dg/cpp2a/array-conv4.C: New test.
* g++.dg/cpp2a/array-conv5.C: New test.
* g++.dg/cpp2a/array-conv6.C: New test.
* g++.dg/cpp2a/array-conv7.C: New test.
* g++.dg/cpp2a/array-conv8.C: New test.
* g++.dg/cpp2a/array-conv9.C: New test.
* g++.old-deja/g++.bugs/900321_01.C: Adjust dg-error.

* testsuite/23_containers/span/lwg3255.cc: Adjust test to match the
post-P0388R4 behavior.

From-SVN: r276771

Fix typo in test name.

From-SVN: r276770

[Darwin] Improve Objective-C NeXT ABI version check.

We were missing a check for the case that user's ABI was > 2 and the
codegen was for 64 bit.

gcc/ChangeLog:

2019-10-09  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (darwin_override_options): Make the check for
Objective-C ABI version more specific for 64bit code.

From-SVN: r276768

[Darwin, machopic 3/n] Set a SYMBOL flag for indirections.

We are able to treat these specially where needed in legitimate address
tests (specifically, they are guaranteed to be pointer-aligned).

gcc/ChangeLog:

2019-10-09  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (machopic_indirect_data_reference): Set flag to
indicate that the new symbol is an indirection.
(machopic_indirect_call_target): Likewise.
* config/darwin.h (MACHO_SYMBOL_FLAG_INDIRECTION): New.
(MACHO_SYMBOL_INDIRECTION_P): New.
(MACHO_SYMBOL_FLAG_STATIC): Adjust bit number.

From-SVN: r276767

PR c++/92032 - DR 1601: Promotion of enum with fixed underlying type.

I've been messing with compare_ics recently and noticed that we don't
implement CWG 1601, which should be fairly easy.  Thus this patch.

The motivating example is
  enum E : char { e };
  void f(char);
  void f(int);
  void g() {
    f(e);
  }
where the call to f was ambiguous but we should choose f(char).

Currently we give f(int) cr_promotion in standard_conversion, while
f(char) remains cr_std, which is worse than cr_promotion.  So I thought
I'd give it cr_promotion also and then add a tiebreaker to compare_ics.

* call.c (standard_conversion): When converting an enumeration with
a fixed underlying type to the underlying type, give it the cr_promotion
rank.
(compare_ics): Implement a tiebreaker as per CWG 1601.

* g++.dg/cpp0x/scoped_enum10.C: New test.
* g++.dg/cpp0x/scoped_enum11.C: New test.

From-SVN: r276766

Update the concepts implementation to conform to C++20.

gcc/c-family/
* c-cppbuiltin.c (c_cpp_builtins): Use new feature test values for
concepts when -std=c++2a. Bump __cpp_concepts to 201907.
* c.opt: Add -Wconcepts-ts.
* c-opts.c (c_common_post_options): Warn when -fconcepts is used
with -std=c++2a. Disable warning for -fconcepts in C++20 mode.
(set_std_cxx2a): Enable concepts by default.
gcc/cp/
* call.c (build_new_function_call): Don't evaluate concepts here.
(constraint_failure): Don't record the template.
(print_z_candidate): Don't extract the template.
* class.c (add_method): When overloading, hide ineligible special
member fns.
(check_methods): Set TYPE_HAS_COMPLEX_* here.
* constexpr.c (cxx_eval_constant_expression): Evaluate concepts.
(maybe_initialize_fundef_copies_table): Remove.
(get_fundef_copy): Use hash_map_safe_get_or_insert.
(clear_cv_and_fold_caches): Clear the satisfaction cache.
* constraint.cc (known_non_bool_p): New.
(parsing_constraint_expression_sentinel): Renamed from
expanding_constraint_sentinel.
(check_constraint_operands): New.
(check_constraint_atom): New.
(finish_constraint_binary_op): New.
(finish_constraint_or_expr): Likewise.
(finish_constraint_and_expr): Likewise.
(finish_constraint_primary_expr): Likewise.
(combine_constraint_expressions): New.
(finish_requires_expr): Add location parm.
(get_concept_definition): Return the initializer of concept definitions.
(get_template_head_requirements): New.
(get_trailing_function_requirements): New.
(deduce_constrained_parameter): Check if the identifier or template-id
is a concept definition.
(resolve_concept_definition_check): Removed.
(resolve_variable_concept_check): Removed.
(resolve_concept_check): New.
(resolve_constraint_check): Handle concept definitions.
converting arguments.
(function_concept_check_p): Removed.
(variable_concept_check_p): Removed.
(unpack_concept_check): New.
(get_concept_check_template): New.
(build_call_check): Moved and renamed to build_function_check.
(build_concept_check_arguments): make static.
(build_function_check): Always do overload resolution
in order to force conversion of template arguments (i.e., actually
check that the use of a concept is valid).
(build_standard_check): Renamed from build_real_concept_check.
(build_real_concept_check): Build checks for C++2a concepts by
(build_wildcard_concept_check): New.
(build_concept_check): Use build_real_concept_check. New overload.
(build_constraints): Save expressions, not normalized constraints.
(build_concept_id): New. Pass tf_warning_or_error.
(build_type_constraint): New.
(finish_type_constraints): New.
(associate_classtype_constraints): Also add constraints to union
types. Note the original declaration in errors. Don't return
error_mark_node in order to avoid an assertion later.
(push_down_pack_expansion): Remove.
(finish_shorthand_constraint): Make fold expressions, not naked
parameter packs. Always apply the constraint to each template argument.
(check_introduction_list): New. Fail if not enough
names are introduced.
(finish_template_introduction): Don't normalize constraints. Pass
tsubst flags. Check for insufficient introductions.
(placeholder_extract_concept_and_args): Handle the template-id case.
Unpack function concept checks correctly.
(tsubst_simple_requirement): Return errors if they occur. Don't
process as a template.
(tsubst_type_requirement): Likewise.
(type_deducible_p): New. Wrap the input expression in parens for the
purpose of deduction.
(expression_convertible_t): New.
(tsubst_compound_requirement): Use new deduction, conversion predicates.
(tsubst_nested_requirement): Return errors if they occur. Don't
process as a template. Instantiate and evaluate the nested requirement.
(tsubst_valid_expression_requirement): New.
(tsubst_simple_requirement): Use tsubst_valid_expression_requirement.
(tsubst_compound_requirement): Use tsubst_valid_expression_requirement.
(check_constaint_variables): New.
(tsubst_constraint_variables): Check that type substitutions are valid.
(tsubst_requires_expr): Likewise. Produce new requires-exprs during
template substitution. Copy the previous local specialization stack,
so references to non-local parameters can be found. Use cp_unevaluated.
(tsubst_constraint): New. Don't evaluate concept checks.
(subst_info): New.
(norm_info): New. Used to build a normalization tree for concept check
diagnostics.
(debug_parameter_mapping): New.
(debug_argument_list): New.
(expand_concept): Removed.
(normalize_logical_operation): Pass subst_info through call.
(normalize_pack_expansion): Remove.
(normalize_simple_requirement): Removed
(normalize_type_requirement): Removed
(normalize_compound_requirement): Removed
(normalize_nested_requirement): Removed
(normalize_requirement): Removed
(normalize_requirements): Removed
(normalize_requires_expression): Removed
(normalize_variable_concept_check): Removed.
(normalize_function_concept_check): Removed.
(normalize_concept_check): Merged all normalize_*_check here.
Substitute through written template arguments before normalizing the
definition. Only substitute the innermost template arguments.
(check_for_logical_overloads): Delete.
(map_arguments): New. Associate template parameters with arguments.
(build_parameter_mapping): New. Extract used parameters.
(normalize_expression): Rewrite.
(normalize_conjunction): Removed
(normalize_disjunction): Removed
(normalize_predicate_constraint): Removed
(normalize_parameterized_constraint): Removed
(normalized_map): New variable.
(get_normalized_constraints): New entry point for normalization.
Establishes a timer.
(get_normalized_constraints_from_info): New.
(get_normalized_constraints_from_decl): New. Turn on template processing
prior to normalization. Handle inheriting ctors. Build the
normalization arguments from the full set of template parameters of the
most general template. This guarantees that we have no concrete arguments
in the parameter mapping (e.g., from template members of class
templates). Cache normalizations.
(normalize_concept_definition): New. Cache normalizations.
(normalize_template_requirements): New.
(normalize_nontemplate_requirements): New.
(normalize_constraint_expression): New.
(tsubst_parameter_mapping): New.
(get_mapped_args): New.
(parameter_mapping_equivalent_p): New. Use template_args_equal.
(atomic_constraints_identical_p): New.
(hash_atomic_constraint): New.
(satisfying_constraint_p): New. Guard against recursive evaluation of
constraints during satisfaction.
(satisfy_conjunction): New.
(satisfy_disjunction): New.
(sat_entry): New class for hashing satisfaction results.
(sat_hasher): New hash traits.
(sat_cache): New.
(get_satisfaction): New. Returns cached satisfaction result.
(save_satisfaction): New. Caches a satisfaction result.
(clear_satisfaction_cache): New.
(satisfaction_cache): New. Helps manage satisfaction cache requests.
(decl_satisfied_cache): New.
(satisfy_atom): New.
(satisfy_constraint_r): New.
(satisfy_constraint): Use new satisfaction algorithm.
(evaluate_concept_check): New.
(evaluate_concept): Removed.
(evaluate_function_concept): Removed.
(evaluate_variable_concept): Removed.
(satisfy_constraint_expression): New.
(constraint_expression_satisfied_p): New.
(constraints_satisfied_p): Use strip_inheriting_ctors. Use
push_/pop_access_scope.
(more_constrained): Normalize before calling out to subsumption. Allow
classes as arguments.
(strictly_subsumes): Allow non-templates as arguments. Accept a new
template argument.
(weakly_subsumes): New.
(at_least_as_constrained): Removed.
(diagnose_other_expression): Removed.
(diagnose_predicate_constraint): Removed.
(diagnose_pack_expansion): Removed.
(diagnose_check_constraint): Removed.
(diagnose_logical_constraint): Removed.
(diagnose_expression_constraint): Removed.
(diagnose_type_constraint): Removed.
(diagnose_implicit_conversion_constraint): Removed.
(diagnose_argument_deduction_constraint): Removed.
(diagnose_exception_constraint): Removed.
(diagnose_parameterized_constraint): Removed.
(diagnose_argument_deduction_constraint): Removed.
(diagnose_argument_deduction_constraint): Removed.
(diagnose_argument_deduction_constraint): Removed.
(diagnose_trait_expr): New.
(diagnose_requires_expr): New.
(diagnose_atomic_constraint): New.
(diagnose_valid_expression) Stop wrongly diagnosing valid expressions.
Don't substitute as if in template decls. This causes substitution
to generate expressions that aren't suitable for use with the noexcept
routines.
(diagnose_valid_type) Likewise.
(diagnose_compound_requirement) Actually emit diagnostics for
the causes of errors.Call force_paren_expr_uneval.
(diagnose_declaration_constraints): Turn on template processing to
suppress certain analyses.
* cp-objcp-common.c (cp_common_init_ts): Make concepts typed.
(cp_get_debug_type): Use hash_map_safe_*.
* cp-tree.h: New function declarations for semantic actions, other
facilities. Remove declaration no longer used or needed. Remove
unused _CONSTR macros.
(LANG_DECL_HAS_MIN): Add CONCEPT_DECL.
(template_info_decl_check): Factor macro check into an inline function.
(DECL_TEMPLATE_INFO): Use new check facility.
(finish_concept_definition): New. Don't invalid concept declarations
with invalid initializers.
(find_template_parameters): New.
(concept_definition_p): New.
(concept_check_p): New.
(variable_concept_check_p): New.
(force_paren_expr_uneval): New.
(ovl_iterator::using_p): A USING_DECL by itself was also
introduced by a using-declaration.
(struct tree_template_info): Use tree_base instead of
tree_common. Add tmpl and args fields.
(TI_TEMPLATE, TI_ARGS): Adjust.
(DECLTYPE_FOR_INIT_CAPTURE): Remove.
(CONSTR_CHECK, CONSTR_INFO, CONSTR_EXPR, CONSTR_CONTEXT): New.
(ATOMIC_CONSTR_MAP, TRAIT_EXPR_LOCATION): New.
(struct tree_trait_expr): Add locus field.
(enum tsubst_flags): Add tf_norm as a hint to generate normalization
context when diagnosing constraint failure.
* cp-tree.def: Remove unused _CONSTR nodes and rename PRED_CONSTR
to ATOMIC_CONSTR.
(CONCEPT_DECL): New.
* cxx-pretty-print.c: Remove constraint printing code.
(pp_cxx_concept_definition): New.
(pp_cxx_template_declaration): Print concept definitions.
(pp_cxx_check_constraint): Update printing for concept definitions.
(pp_cxx_nested_name_specifier): Fix a weird
case where we're printing '::::' for concepts.
(simple_type_specifier): Print requirements for placeholder types.
(pp_cxx_constrained_type_spec): Print the associated requirements of
a placeholder type.
(pp_cxx_compound_requirement): Add space before the '->'.
(pp_cxx_parameter_mapping): Print the parameter mapping.
(pp_cxx_atomic_constraint): Use the function above.
* decl.c (redeclaration_error_message): New error for concepts.
(grokdeclarator): Check for and disallow decltype(auto) in parameter
declarations.
(grokfndecl): Don't normalize constraints. Add check for constraints
on declaration.
(grokvardecl): Don't normalize constraints.
(grok_special_member_properties): Don't set TYPE_HAS_COMPLEX_*.
(function_requirements_equivalent_p): New. Compare trailing
requires clauses. Compare combined constraints in pre-C++20 mode.
(decls_match): Compare trailing requires clauses. Compare template
heads for function templates. Remove old constraint comparison.
Simplify comparison of functions, function templates.
(duplicate_function_template_decls): New. Refactor a nasty if
condition into a single predicate.
(require_deduced_type): Don't complain if we already complained about
deduction failure.
(finish_function): Perform auto deduction to ensure that constraints
are checked even when functions contain no return statements. Only do
auto deduction if we haven't previously seen any return statements.
This prevents multiple diagnostics of the same error.
(store_decomp_type): Remove.
(cp_finish_decomp): Use hash_map_safe_put.
* error.c: Remove constraint printing code.
(dump_decl): Dump concept definitions. Handle wildcard declarations.
(dump_template_decl): Likewise.
(dump_type): Print associated requirements for placeholder
types.
(rebuild_concept_check): New.
(maybe_print_single_constraint_context): New.
(maybe_print_constraint_context): Recursively print nested contexts.
* init.c (get_nsdmi): Use hash_map_safe_*.
* lambda.c (maybe_add_lambda_conv_op): Bail if deduction failed.
(add_capture): Copy parameter packs from init.
(lambda_capture_field_type): Always use auto for init-capture.
* logic.cc: Completely rewrite.
(constraint_hash): New.
(clause/ctor): Save atoms in the hash table.
(replace): Save atoms during replacement.
(insert): Save atoms during insertion.
(contains): Only search the hash table for containment.
(clause): Keep a hash of atomic constraints.
(clause::clause): Explicitly copy the hash table when copying.
(disjunction_p, conjunction_p, atomic_p, dnf_size, cnf_size): New.
(diagnose_constraint_size): New.
(subsumes_constraints_nonnull): Compare the sizes of normalized formula
to determine the cheapest decomposition.
* name-lookup.c (diagnose_name_conflict): Diagnose name issues with
concepts.
(matching_fn_p): Check constraints.
(push_class_level_binding_1): Move overloaded functions case down,
accept FUNCTION_DECL as target_decl.
* parser.c (enum required_token): New required token for auto.
(make_location): Add overload taking lexer as last parm.
(cp_parser_required_error): Diagnose missing auto.
(cp_parser_diagnose_ungrouped_constraint_plain): New.
(cp_parser_diagnose_ungrouped_constraint_plain): New.
(cp_parser_constraint_primary_expression): New. Tentatively parse the
primary expression. If that fails tentatively parse a lower
precedence expression in order to diagnose the error.
(cp_parser_check_non_logical_constraint): New. Performs a trial
parse of the right-hand-side of non-logical operators in order to
generate good diagnostics.
(cp_parser_constraint_logical_and_expression): New.
(cp_parser_constraint_logical_or_expression): New.
(cp_parser_requires_clause_expression): New.
(cp_parser_requires_clause): Renamed to cp_parser_constraint_expression.
(cp_parser_requires_clause_opt): Parse the requires-clause differently
in -fconcepts and -std=c++2a modes.
(cp_parser_requirement_list): Rename to cp_parser_requirement_seq.
Rewrite so that semicolons are parsed
along with requirements, not the sequence.
(cp_parser_simple_requirement): Expect a semicolon at end.
(cp_parser_compound_requirement): Expect a semicolon at end. Only
allow trailing-return-type with -fconcepts-ts.
(cp_parser_nested_requirement): Expect a semicolon at end. Parse
constraint-expressions.
(cp_parser_concept_definition): New. Don't fail parsing the concept
definition if the initializer is ill-formed. Don't declare the concept
before parsing the initializer.
(cp_parser_constraint_expression): Declare earlier.
(cp_parser_type_requirement): Current scope is not valid.
(cp_parser_requires_expression): Commit to the tentative parse.
(cp_parser_decl_specifier_seq): Warn when concept appears to be used
as a decl-specifier.
(cp_parser_template_declaration_after_parameters): Parse concept
definitions.
(cp_parser_template_id): Don't try to resolve a concept template-id yet.
(cp_parser_template_id_expr): Resolve it as a concept check.
(cp_parser_decl_specifier_seq): Warn on 'concept bool'.
(cp_parser_type_parameter): Combine expressions not
constraints.
(cp_parser_explicit_template_declaration): Combine expressions not
constraints.
(cp_parser_maybe_concept_name): Removed.
(cp_parser_simple_type_specifier): Handle an error condition of
a bad constrained type specifier. Expect auto or decltype after
a concept name. Also handle the case where we have a template-id
as a concept check.
(cp_parser_template_introduction): Diagnose errors on invalid
introductions. Give up if it doesn't start with a concept name.
Pedwarn if not -fconcepts-ts.
(synthesize_implicit_template_parm): Don't do consistent binding.
Use a new flag for constrained parameters. Combine expressions,
not constraints. Fail if we get a placeholder in block scope.
Placeholders that do not constrain types are not allowed in parameter
declarations, so don't handle them.
(cp_parser_placeholder_type_specifier): New. Implement parsing of
placeholder type specifiers following a concept name or partial
concept check. Disallow decltype(auto) parameters.
(cp_parser_nested_name_specifier_opt): If the token is already
CPP_NESTED_NAME_SPECIFIER, leave it alone.
(cp_parser_id_expression, cp_parser_unqualified_id): Call
cp_parser_template_id_expr.
(cp_parser_placeholder_type_specifier): Add tentative parm.  Don't
expect a WILDCARD_DECL.
(cp_parser_trait_expr): Pass trait_loc down.
(cp_parser_postfix_expression): Do set location of dependent member
call.
* pt.c (finish_concept_definition): New.
(push_template_decl_real): Handle concept definitions.
(start_concept_definition): Let push_template_decl_real handle the
creation of the template.
(get_constraints): Return null if the table hasn't been initialized.
(tsubst_copy_and_build): Build template-id expressions for concept
checks.
[TRAIT_EXPR]: Pass trait_loc down.
(lookup_template_class_1): Add the template name to the constraint
failure diagnostic.
(lookup_and_finish_template_variable): Build concept checks
with the correct arguments.
(tsubst_function_decl): Don't substitute through constraints.
Always associate constraints with functions.
(template_parm_level_and_index): Make non-static.
(for_each_template_parm_r): Handle requires expressions.
(keep_template_parm): New.
(find_template_parameters): New.
(more_specialized_fn): Change how winners and losers are chosen.
(make_constrained_auto): Don't normalize constraints.
(template_parameters_equivalent_p): New. Compare template
parameters. Add a comparison for implicitly vs. explicitly declared
parameters.
(template_parameter_lists_equivalent_p): New. Compare template
parameter lists.
(template_requirements_equivalent_p): New.
(template_heads_equivalent_p): New. Compare template heads.
(template_parameter_constraints_equivalent_p): New.
(is_compatible_template_arg): Use weakly_subsumes.
(maybe_new_partial_specialization): Use new constraint comparison
for finding specializations.
(process_partial_specialization): Pass main template as argument.
(more_specialized_partial_spec): Don't immediately return when
detecting a winner.
(make_constrained_auto): Handle concept definitions.
(do_auto_deduction): Update auto deduction for new concept model.
Extract the function concept correctly; rename constr to check to
reflect the kind of node.
(tsubst): Adjust wildcard argument during substitution.
[DECLTYPE_TYPE]: Remove init-capture handling.
(tsubst_copy_and_build): Build concept checks, not template ids.
Defer checks of function concepts. Handle concepts before variable
templates. Handle calls to function concepts explicitly.
(coerce_template_parms): Use concept_definition_p. Handle a deduction
error where a potentially empty pack can be supplied after the last
parameter of a concept.
(finish_template_variable): Don't process concepts here.
(instantiation_dependent_r): Use concept_check_p.
(tsubst_template_args): Make non-static.
(make_constrained_placeholder_type): New. Refactored from
make_constrained_auto.
(make_constrained_auto) Use make_constrained_placeholder_type.
(make_constrained_decltype_auto) New.
(tsubst_function_parms): New.
(value_dependent_expression_p) [TEMPLATE_ID_EXPR]: Use
concept_definition_p.
(push_access_scope, pop_access_scope): No longer static.
(tsubst_template_parm): Substitute TEMPLATE_PARM_CONSTRAINTS.
(tsubst_friend_function): Use tsubst_constraint. Use generic_targs_for.
(get_underlying_template) Use generic_targs_for.
(uses_parameter_packs): Return tree.
(gen_elem_of_pack_expansion_instantiation): Don't push
local_specialization_stack.
(prepend_one_capture): New.
(tsubst_lambda_expr): Use prepend_one_capture. Don't touch
local_specializations.
(template_parms_level_to_args): No longer static.
(add_outermost_template_args): Likewise.
(find_template_parameter_info): New. Provide context for finding
template parameters.
(keep_template_parm): Don't keep parameters declared at depth levels
greater than those of the template parameters of the source declaration.
Don't propagate cv-qualified types. Return 0, so we find all template
parameters, not the just first.
(any_template_parm_r): New. Handle cases that are mishandled by
for_each_template_parm_r.
(generic_targs_for): Factor out of coerce_template_args_for_ttp.
(tsubst_argument_pack): Factor out of tsubst_template_args.
(constraint_sat_entry): Removed.
(constraint_sat_hasher): Removed.
(concept_spec_entry): Removed.
(concept_spec_hasher): Removed.
(constraint_memos): Removed.
(concept_memos): Removed.
(lookup_constraint_satisfaction): Removed.
(memoize_constraint_satisfaction): Removed.
(lookup_concept_satisfaction): Removed.
(memoize_concept_satisfaction): Removed.
(concept_expansions): Removed.
(get_concept_expansion): Removed.
(save_concept_expansion): Removed.
(init_constraint_processing): Remove initialization of non-existing
resources.
(find_template_requirement): New. Search for the sub-requirement
within the associated constraints.
(convert_generic_types_to_packs): Also transform the associated
constraint and update the current template requirements.
(store_defaulted_ttp, lookup_defaulted_ttp): Remove.
(add_defaults_to_ttp): Use hash_map_safe_*.
* semantics.c (finish_call_expr): Diagnose calls to concepts.
Handle concept checks explicitly.
(finish_id_expression): Evaluate variable concepts as part of
id-expression processing. Don't treat variable concepts as variables,
and don't process function concepts as plain id-expressions.
(force_paren_expr): Add even_uneval parm.
(finish_trait_expr): Add location parm.
* tree.c (special_memfn_p): New.
(cp_expr_location): Handle TRAIT_EXPR.
* typeck.c (check_return_expr): Actually use the diagnostic kind
when performing return-type deduction.
* typeck2.c (build_functional_cast): Don't rely on the location of
'auto'.
gcc/testsuite/
* lib/target-supports.exp (check_effective_target_concepts): Check
for std=c++2a.
gcc/
* doc/invoke.texi: Document -fconcepts-ts.

From-SVN: r276764

decl.c (grok_ctor_properties): Use DECL_SOURCE_LOCATION.

/cp
2019-10-09  Paolo Carlini  <paolo.carlini@oracle.com>

* decl.c (grok_ctor_properties): Use DECL_SOURCE_LOCATION.
* typeck.c (cp_build_binary_op): Use the op_location_t argument
in many error messages.

/testsuite
2019-10-09  Paolo Carlini  <paolo.carlini@oracle.com>

* c-c++-common/Waddress-1.c: Test locations too.
* c-c++-common/Wpointer-compare-1.c: Likewise.
* c-c++-common/Wshift-count-negative-1.c: Likewise.
* c-c++-common/Wshift-count-overflow-1.c: Likewise.
* c-c++-common/Wshift-negative-value-1.c: Likewise.
* c-c++-common/Wshift-negative-value-2.c: Likewise.
* c-c++-common/Wshift-negative-value-5.c: Likewise.
* c-c++-common/pr48418.c: Likewise.
* c-c++-common/pr65830.c: Likewise.
* c-c++-common/pr69764.c: Likewise.
* g++.dg/cpp0x/constexpr-array-ptr10.C: Likewise.
* g++.dg/cpp0x/nullptr37.C: Likewise.
* g++.dg/template/crash126.C: Likewise.
* g++.dg/template/crash129.C: Likewise.
* g++.dg/warn/Wextra-3.C: Likewise.
* g++.dg/warn/Wfloat-equal-1.C: Likewise.
* g++.dg/warn/Wstring-literal-comparison-1.C: Likewise.
* g++.dg/warn/Wstring-literal-comparison-2.C: Likewise.
* g++.dg/warn/pointer-integer-comparison.C: Likewise.
* g++.old-deja/g++.jason/crash8.C: Likewise.

From-SVN: r276763

PR libstdc++/91057 set locale::id::_M_index atomically

If two threads see _M_index==0 concurrently they will both try to set
it, potentially storing the facet at two different indices in the array.

Either set the _M_index data member using an atomic compare-exchange
operation or while holding a mutex.

Also move the LONG_DOUBLE_COMPAT code into a separate function to remove
the visual noise it creates.

PR libstdc++/91057
* src/c++98/locale.cc (locale::id::_M_id()) [__GTHREADS]: Use atomic
compare-exchange or double-checked lock to ensure only one thread sets
the _M_index variable.
[_GLIBCXX_LONG_DOUBLE_COMPAT]: Call find_ldbl_sync_facet to detect
facets that share another facet's ID.
[_GLIBCXX_LONG_DOUBLE_COMPAT] (find_ldbl_sync_facet): New function.

From-SVN: r276762

tree-vect-loop.c (vect_is_simple_reduction): Simplify and allow stmts other than GIMPLE_ASSIGN in nested cycles.

2019-10-09  Richard Biener  <rguenther@suse.de>

* tree-vect-loop.c (vect_is_simple_reduction): Simplify and
allow stmts other than GIMPLE_ASSIGN in nested cycles.

* gcc.dg/vect/vect-outer-call-1.c: New testcase.

From-SVN: r276760

PR libstdc++/78552 only construct std::locale for C locale once

PR libstdc++/78552
* src/c++98/locale_init.cc (locale::classic()): Do not construct a new
locale object for every call.
(locale::_S_initialize_once()): Construct C locale here.

From-SVN: r276758

[PR92036] Add 'libgomp.oacc-c-c++-common/data-firstprivate-1.c'

libgomp/
PR middle-end/92036
* testsuite/libgomp.oacc-c-c++-common/data-firstprivate-1.c: New
file.

From-SVN: r276757

avr-mmcu.texi: Re-generate because config/avr/avr-devices.c was changed in r269487.

* doc/avr-mmcu.texi: Re-generate because config/avr/avr-devices.c
was changed in r269487.

From-SVN: r276756

* config/avr/avr.md: Fix typo in a comment.

From-SVN: r276755

re PR testsuite/91884 (libgomp testsuite: (not) using a specific driver for C++, Fortran)

2019-10-09  Tobias Burnus  <tobias@codesourcery.com>

        PR testsuite/91884
        * testsuite/libgomp.fortran/fortran.exp: Conditionally
        add -lquadmath.
        * testsuite/libgomp.oacc-fortran/fortran.exp: Ditto.

From-SVN: r276754

re PR libgomp/92028 (OpenACC 'host_data' execution test regressions with nvptx offloading)

PR libgomp/92028
* target.c (gomp_map_vars_internal): Readd the previous
GOMP_MAP_USE_DEVICE_PTR handling code in the first loop,
though do that just in the !not_found_cnt case.

From-SVN: r276753

tree-vectorizer.h (_stmt_vec_info::reduc_vectype_in): New.

2019-10-08  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (_stmt_vec_info::reduc_vectype_in): New.
(_stmt_vec_info::force_single_cycle): Likewise.
(STMT_VINFO_FORCE_SINGLE_CYCLE): New.
(STMT_VINFO_REDUC_VECTYPE_IN): Likewise.
* tree-vect-loop.c (vectorizable_reduction): Set
STMT_VINFO_REDUC_VECTYPE_IN and STMT_VINFO_FORCE_SINGLE_CYCLE.
(vect_transform_reduction): Use them to remove redundant code.
(vect_transform_cycle_phi): Likewise.

From-SVN: r276752

use call-clobbered reg to disalign the stack

Some x86 tests of stack realignment, that disaligned the stack with
pushes and pops, failed when the compiler was configured to tune for a
target that preferred to accumulate outgoing arguments: the stack
space is reserved before the asm push, the call sequence overwrites
the saved register, and then the asm pop restores the overwritten
value.  Since that's a call-preserved register in 32-bit mode, it
should be preserved unchanged, but isn't.

Merely changing the register to a call-clobbered one would be enough,
but the tests would remain fragile and prone to failure due to other
optimizations, so I arranged for the compiler to be made aware of the
register used for the push and the pop, so it won't use it for
something else, and forced the function to use a frame pointer, so
that it won't use stack pointer offsets for local variables: the
offsets would likely be wrong between the asm push and pop.

for  gcc/testsuite/ChangeLog

* gcc.target/i386/20060512-1.c (sse2_test): Use a
call-clobbered register variable for stack-disaligning push
and pop.  Require a frame pointer.
* gcc.target/i386/20060512-3.c (sse2_test): Likewise.

From-SVN: r276751

Daily bump.

From-SVN: r276750

PR c++/92001 - missing -Wclass-memaccess with array as first argument to memset

gcc/cp/ChangeLog:

PR c++/92001
* call.c (maybe_warn_class_memaccess): Handle arrays.

gcc/testsuite/ChangeLog:

PR c++/92001
* g++.dg/Wclass-memaccess-5.C: New test.

From-SVN: r276725

re PR tree-optimization/90836 (Missing popcount pattern matching)

2019-10-08  Dmitrij Pochepko <dmitrij.pochepko@bell-sw.com>

PR tree-optimization/90836
* lib/target-supports.exp (check_effective_target_popcount)
(check_effective_target_popcountll): New effective targets.
* gcc.dg/tree-ssa/popcount4.c: New test.
* gcc.dg/tree-ssa/popcount4l.c: New test.
* gcc.dg/tree-ssa/popcount4ll.c: New test.

From-SVN: r276722

re PR tree-optimization/90836 (Missing popcount pattern matching)

2019-10-08  Dmitrij Pochepko <dmitrij.pochepko@bell-sw.com>

PR tree-optimization/90836
* gcc/match.pd (popcount): New pattern.

From-SVN: r276721

PR middle-end/92026 - gcc.dg/Wstringop-overflow-18.c FAIL

PR middle-end/92026 - gcc.dg/Wstringop-overflow-18.c FAIL
PR middle-end/92014 - bogus warning: writing 8 bytes into a region of size 1 in timezone/zic.c

gcc/ChangeLog:
* tree-ssa-strlen.c (count_nonzero_bytes): Avoid recursing for MEM_REF
again once nbytes has been set.  Set the access size when not yet set.

gcc/testsuite/ChangeLog:

PR middle-end/92014
* gcc.dg/Wstringop-overflow-19.c: New test.

From-SVN: r276711

[Darwin] Remove code deprecated in 4.x.

This removes some code that should be dead.
Given no reported problems from the warning since 4.6 this seems reasonable.

gcc/ChangeLog:

2019-10-08  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (machopic_select_section): Remove dead code for
old Objective-C section selection method, replace with unreachable.

From-SVN: r276709

[Darwin, machopic 2/n] Compute and cache indirection rules.

This caches a check for the requirement to indirect a symbol in the Darwin
ABI, and uses it where needed.  We also ensure that we place the indirection
pointers into the non-lazy symbol pointers section.  Other placements have
occurred with various platform toolchains - but these seem to have been
unintentional so we match current platform toolchains.

gcc/ChangeLog:

2019-10-08  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (machopic_indirect_data_reference): Check for
required indirections before making direct access to defined
values.
(machopic_output_indirection): Place the indirected pointes for
required indirections into the non-lazy symbol pointers section.
(darwin_encode_section_info):
* config/darwin.h (MACHO_SYMBOL_FLAG_MUST_INDIRECT): New.
(MACHO_SYMBOL_MUST_INDIRECT_P): New.

From-SVN: r276708

re PR rtl-optimization/91994 (r276327 breaks -mvzeroupper)

PR target/91994
* config/i386/i386.c (x86_avx_u128_mode_needed): Use SSE_REG
instead of ALL_SSE_REG to check if function call preserves some
256-bit SSE registers.

From-SVN: r276707

re PR fortran/91801 (ICE in gfc_simplify_reshape, at fortran/simplify.c:6733)

2019-10-08  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91801
* simplify.c (gfc_simplify_reshape): Convert a gcc_assert into a
gfc_error as a user can easily hit the condition.

2019-10-08  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91801
* gfortran.dg/pr91801.f90: New test.

From-SVN: r276706

DR 685 - Integral promotion of enum ignores fixed underlying type.

* g++.dg/cpp0x/scoped_enum9.C: New test.

From-SVN: r276705

re PR testsuite/92025 (gcc.dg/Wstringop-overflow-12.c XPASSes)

gcc/testsuite/ChangeLog:

PR tetsuite/92025
* gcc.dg/Wstringop-overflow-12.c: Remove xfail.

2019-10-08  Richard Biener  <rguenther@suse.de>

From-SVN: r276703

Change the library search path when using --with-advance-toolchain

From: Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com>

Remove all -L directories from LINK_OS_EXTRA_SPEC32 and
LINK_OS_EXTRA_SPEC64 so that user directories specified at
build time have higher preference over the advance toolchain libraries.

Set MD_STARTFILE_PREFIX to $prefix/lib/ and MD_STARTFILE_PREFIX_1 to
$at/lib/ so that a compiler library has preference over the Advance
Toolchain libraries.

* config.gcc: Move -L usage from LINK_OS_EXTRA_SPEC32 and
LINK_OS_EXTRA_SPEC64 to MD_STARTFILE_PREFIX and
MD_STARTFILE_PREFIX_1 when using --with-advance-toolchain.

From-SVN: r276702

tree-vectorizer.h (_stmt_vec_info::v_reduc_type): Remove.

2019-10-08  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (_stmt_vec_info::v_reduc_type): Remove.
(_stmt_vec_info::is_reduc_info): Add.
(STMT_VINFO_VEC_REDUCTION_TYPE): Remove.
(vectorizable_condition): Remove.
(vectorizable_shift): Likewise.
(vectorizable_reduction): Adjust.
(info_for_reduction): New.
* tree-vect-loop.c (vect_force_simple_reduction): Fold into...
(vect_analyze_scalar_cycles_1): ... here.
(vect_analyze_loop_operations): Adjust.
(needs_fold_left_reduction_p): Simplify for single caller.
(vect_is_simple_reduction): Likewise.  Remove stmt restriction
for nested cycles not part of double reductions.
(vect_model_reduction_cost): Pass in the reduction type.
(info_for_reduction): New function.
(vect_create_epilog_for_reduction): Use it, access reduction
meta off the stmt info it returns.  Use STMT_VINFO_REDUC_TYPE
instead of STMT_VINFO_VEC_REDUCTION_TYPE.
(vectorize_fold_left_reduction): Remove pointless assert.
(vectorizable_reduction): Analyze the full reduction when
visiting the outermost PHI.  Simplify.  Use STMT_VINFO_REDUC_TYPE
instead of STMT_VINFO_VEC_REDUCTION_TYPE.  Direct reduction
stmt code-generation to vectorizable_* in most cases.  Verify
code-generation only for cases handled by
vect_transform_reductuon.
(vect_transform_reduction): Use info_for_reduction to get at
reduction meta.  Simplify.
(vect_transform_cycle_phi): Likewise.
(vectorizable_live_operation): Likewise.
* tree-vect-patterns.c (vect_reassociating_reduction_p): Look
at the PHI node for STMT_VINFO_REDUC_TYPE.
* tree-vect-slp.c (vect_schedule_slp_instance): Remove no
longer necessary code.
* tree-vect-stmts.c (vectorizable_shift): Make static again.
(vectorizable_condition): Likewise.  Get at reduction related
info via info_for_reduction.
(vect_analyze_stmt): Adjust.
(vect_transform_stmt): Likewise.
* tree-vectorizer.c (vec_info::new_stmt_vec_info): Initialize
STMT_VINFO_REDUC_TYPE instead of STMT_VINFO_VEC_REDUCTION_TYPE.

* gcc.dg/vect/pr65947-1.c: Adjust.
* gcc.dg/vect/pr65947-13.c: Likewise.
* gcc.dg/vect/pr65947-14.c: Likewise.
* gcc.dg/vect/pr65947-4.c: Likewise.
* gcc.dg/vect/pr80631-1.c: Likewise.
* gcc.dg/vect/pr80631-2.c: Likewise.

From-SVN: r276700

Fortran - fix OpenMP 'target simd'

gcc/fortran/
* parse.c (parse_executable): Add missing ST_OMP_TARGET_SIMD.

libgomp/
* testsuite/libgomp.fortran/target-simd.f90: New.

From-SVN: r276698

Add makefile target to update HTML files in source tree

Also remove the creation of the html/ext sub-directory, which has been
unused since revision r245258.

* doc/Makefile.am (doc-html-docbook-regenerate): New target.
(${docbook_outdir}/html): Do not create unused 'html/ext' directory.
* doc/Makefile.in: Regenerate.
* doc/xml/manual/documentation_hacking.xml: Document new target.
* doc/html/*: Regenerate.

From-SVN: r276695

Fortran - Improve OpenMP/OpenACC diagnostic

gcc/fortran/
* match.h (gfc_match_omp_eos_error): Renamed from gfc_match_omp_eos.
* openmp.c (gfc_match_omp_eos): Make static.
(gfc_match_omp_eos_error): New.
* parse.c (matchs, matchdo, matchds): Do as done for 'matcho' -
        if error occurred after OpenMP/OpenACC directive matched, do not
try other directives.
(decode_oacc_directive, decode_omp_directive): Call new function
instead.

testsuite/
* gfortran.dg/goacc/continuation-free-form.f95: Update dg-error.

From-SVN: r276694

Restore URL for Austern article on allocators

This reverts "Remove broken URL from libstdc++ manual" by restoring the
link, but using an archived copy from the Wayback Machine.

* doc/xml/manual/allocator.xml: Use archived copy of CUJ article.
* doc/html/*: Regenerate.

From-SVN: r276693

Remove Cell Broadband Engine SPU targets

Follow-up to trunk 275343:

* MAINTAINERS: Add back Trevor Smigiel; move into Write After
Approval section.

From-SVN: r276692

Extend 'libgfortran/runtime/minimal.c' per r274599 "PR fortran/68401 Improve allocation error message"

libgfortran/
PR fortran/68401
* runtime/minimal.c (os_error_at): New function.

From-SVN: r276691

Revise 'libgfortran/runtime/minimal.c' to better conform to the original sources

libgfortran/
* runtime/minimal.c: Revise.

From-SVN: r276690

Remove '>>>' merge marker from changelog

From-SVN: r276689

[AArch64] Limit simd-abi-9.c function body test to LP64

On ILP32 targets, there's an extra UXTW instruction to extend
the incoming pointer before the load.  It doesn't seem worth
complicating the test for that, since all we're checking is
that an optimisation takes place, and that optimisation isn't
related to pointer size.

2019-10-08  Richard Sandiford  <richard.sandiford@arm.com>

gcc/testsuite/
* gcc.target/aarch64/torture/simd-abi-9.c: Require LP64 for
the function body test.

From-SVN: r276688

Make C2X imply -fno-fp-int-builtin-inexact.

Since TS 18661-1 has been integrated into C2X, this patch makes C2X
imply -fno-fp-int-builtin-inexact.

Bootstrapped with no regressions on x86_64-pc-linux-gnu.

gcc:
* doc/invoke.texi (-ffp-int-builtin-inexact): Document
-fno-fp-int-builtin-inexact default for C2X.

gcc/c-family:
* c-opts.c (c_common_post_options): Set
-fno-fp-int-builtin-inexact for C2X.

gcc/testsuite:
* gcc.dg/torture/builtin-fp-int-inexact-c2x.c: New test.

From-SVN: r276686

Daily bump.

From-SVN: r276685

re PR tree-optimization/91532 ([SVE] Redundant predicated store in gcc.target/aarch64/fmla_2.c)

2019-10-07  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>
    Richard Biener  <rguenther@suse.de>

PR tree-optimization/91532
* tree-if-conv.c: Include tree-ssa-dse.h.
(ifcvt_local_dce): Change param from bb to loop,
and call dse_classify_store.
(tree_if_conversion): Pass loop instead of loop->header as arg
to ifcvt_local_dce.
* tree-ssa-dse.c: Include tree-ssa-dse.h.
(delete_dead_or_redundant_assignment): Remove static qualifier from
declaration, and add prototype in tree-ssa-dse.h.
(dse_store_status): Move to tree-ssa-dse.h.
(dse_classify_store): Remove static qualifier and add new tree param
stop_at_vuse, and add prototype in tree-ssa-dse.h.
* tree-ssa-dse.h: New header.

Co-Authored-By: Richard Biener <rguenther@suse.de>
From-SVN: r276681

Revert: 2019-10-07  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config/msp430/msp430.md: Revert: Group zero_extend* insns together.

From-SVN: r276680

msp430.md: Group zero_extend* insns together.

2019-10-07  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config/msp430/msp430.md: Group zero_extend* insns together.

From-SVN: r276679

[Darwin, machopic 1/n] Consider visibility in indirections.

For weak, hidden vars the indirection should just be as normal, that
is that the indirections for such symbols should appear in the non-lazy
symbol pointers table, not in the .data section.

gcc/ChangeLog:

2019-10-07  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (machopic_output_indirection): Don't put
hidden symbol indirections into the .data section, use the
non-lazy symbol pointers section as normal.
(darwin_encode_section_info): Record if a symbol is hidden.
* config/darwin.h (MACHO_SYMBOL_FLAG_HIDDEN_VIS): New.
(MACHO_SYMBOL_HIDDEN_VIS_P): New.

From-SVN: r276675

[Darwin, machopic 0/n] Initial tidy of Mach-O symbol handling.

We want to improve the detection and caching of symbol-properties
so that (a) we can make the compiler's output match the platform
norms (b) we can improve efficiency by checking flags instead of
inspecting strings. (c) The fix for PR71767 was a largish hammer
and we want to reduce the number of symbols that are made linker-
visible.

This first patch is largely typographical changes with no functional
difference intended:

- Tries to ensure that there's no overlap between the symbols used in
the Mach-O case and those declared in the i386 or rs6000 port trees.

 - Some improvement to comments.

- Makes the naming of the symbol flags consistent with other uses.

 - Provides a predicate macro for each use.

gcc/ChangeLog:

2019-10-07  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (machopic_symbol_defined_p): Use symbol flag
predicates instead of accessing bits directly.
(machopic_indirect_call_target): Likewise.
(machopic_output_indirection): Likewise.
(darwin_encode_section_info): Improve description.  Use renamed
symbol flags.  Use predicate macros for variables and functions.
* config/darwin.h:
Rename MACHO_SYMBOL_VARIABLE to MACHO_SYMBOL_FLAG_VARIABLE.
Rename MACHO_SYMBOL_DEFINED to MACHO_SYMBOL_FLAG_DEFINED.
Rename MACHO_SYMBOL_STATIC to MACHO_SYMBOL_FLAG_STATIC.
(MACHO_SYMBOL_VARIABLE_P): New.
(MACHO_SYMBOL_DEFINED_P):New.
(MACHO_SYMBOL_STATIC_P): New.
* config/i386/darwin.h (MACHO_SYMBOL_FLAG_VARIABLE): Delete.
(SYMBOL_FLAG_SUBT_DEP): New.
* config/rs6000/darwin.h (SYMBOL_FLAG_SUBT_DEP): New.

From-SVN: r276674

msp430.c (msp430_file_end): s/msp_/msp430_/

2019-10-07  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config/msp430/msp430.c (msp430_file_end): s/msp_/msp430_/
(msp430_expand_epilogue): Likewise.
* config/msp430/predicates.md: Likewise.
* config/msp430/msp430.md: Likewise.
Replace blocks of 8 spaces with tabs.

From-SVN: r276671

msp430-protos.h (msp430_split_addsi): New prototype.

2019-10-07  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config/msp430/msp430-protos.h (msp430_split_addsi): New prototype.
* config/msp430/msp430.c (msp430_split_addsi): New.
* config/msp430/msp430.md: Call msp430_split_addsi () instead of using
a block of C code for splitting addsi.

From-SVN: r276670

i386-expand.c (ix86_expand_floorceildf_32, [...]): Reorder functions.

* config/i386/i386-expand.c (ix86_expand_floorceildf_32,
ix86_expand_rounddf_32): Reorder functions.
* config/i386/i386-protos.h: Update..

From-SVN: r276668

MSP430: Don't generate 430X insns when handling data in the lower memory region

gcc/ChangeLog:

2019-10-07  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config.in: Regenerate.
* config/msp430/constraints.md: Fix docstring for "Ys" constraint.
Add new "Yx" constraint.
* config/msp430/driver-msp430.c (msp430_propagate_region_opt): New spec
function.
* config/msp430/msp430-protos.h (msp430_op_not_in_high_mem): New
prototype.
* config/msp430/msp430.c (msp430_option_override): Allow the lower
code/data region to be selected in the small memory model.
(msp430_section_attr): Don't warn if the "section" and "lower"
attributes are used together.
(msp430_handle_generic_attribute): Likewise.
(msp430_var_in_low_mem): New function.
(TARGET_ENCODE_SECTION_INFO): Define.
(msp430_encode_section_info): New function.
(gen_prefix): Return early in the small memory model.
Require TARGET_USE_LOWER_REGION_PREFIX to be set before adding the
".lower" prefix if -m{code,data}-region=lower have been passed.
(msp430_output_aligned_decl_common): Emit common symbols when
-mdata-region=lower is passed unless TARGET_USE_LOWER_REGION_PREFIX is
set.
(TARGET_ASM_FILE_END): Define.
(msp430_file_end): New function.
(msp430_do_not_relax_short_jumps): Allow relaxation when
function will be in the lower region.
(msp430_op_not_in_high_mem): New function.
(msp430_print_operand): Check "msp430_op_not_in_high_mem" for
the 'X' operand selector.
Clarify comment for 'x' operand selector.
* config/msp430/msp430.h (LINK_SPEC): Propagate
-m{code,data}-region to the linker via spec function
msp430_propagate_region_opt.
(msp430_propagate_region_opt): New prototype.
(EXTRA_SPEC_FUNCTIONS): Add msp430_propagate_region_opt.
(SYMBOL_FLAG_LOW_MEM): Define.
* config/msp430/msp430.md (addsipsi3): Add missing "%X" operand
selector.
(zero_extendqihi2): Fix operand number used by "%X" selector.
(zero_extendqisi2): Likewise.
(zero_extendhisi2): Likewise.
(movqi): Use "Yx" constraint in place of "%X" operand selector.
(movhi): Likewise.
(addqi3): Likewise.
(addhi3): Likewise.
(addsi3): Likewise.
(addhi3_cy): Likewise.
(addchi4_cy): Likewise.
(subqi3): Likewise.
(subhi3): Likewise.
(subsi3): Likewise.
(bic<mode>3): Likewise.
(and<mode>3): Likewise.
(ior<mode>3): Likewise.
(xor<mode>3): Likewise.
(slli_1): Add missing "%X" operand selector.
(slll_1): Likewise.
(slll_2): Likewise.
(srai_1): Likewise.
(sral_1): Likewise.
(sral_2): Likewise.
(srli_1): Likewise.
(srll_1): Likewise.
(cbranchqi4_real): Use "Yx" constraint in place of "%X" operand
selector.
(cbranchhi4_real): Likewise.
(cbranchqi4_reversed): Likewise.
(cbranchhi4_reversed): Likewise.
(*bitbranch<mode>4): Likewise.
(*bitbranch<mode>4_z): Remove unnecessary "%x" operand selector.
* config/msp430/msp430.opt (mcode-region=): Set default to
MSP430_REGION_LOWER. Improve docstring.
(mdata-region=): Likewise.
(muse-lower-region-prefix): New option.
* config/msp430/t-msp430 (MULTILIB_OPTIONS): Add
mdata-region=none multilib.
(MULTILIB_MATCHES): Set mdata-region={upper,either} to match
mdata-region=none multilib.
MULTILIB_EXCEPTIONS: Remove.
MULTILIB_REQUIRED: Define.
* configure: Regenerate.
* configure.ac: Define HAVE_AS_GNU_ATTRIBUTE and
HAVE_AS_MSPABI_ATTRIBUTE if GAS version >= 2.33.50.
* doc/extend.texi: Clarify comment for {upper,lower,either}
function attributes.
Add separate description for "lower" variable attribute.

gcc/testsuite/ChangeLog:

2019-10-07  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* gcc.target/msp430/430x-insns.c: New test.
* gcc.target/msp430/data-attributes-2.c: Remove dg-warning
directives for conflicts between the "section" and "lower" attributes.
* gcc.target/msp430/msp430.exp
(check_effective_target_msp430_region_not_lower): New.
(check_effective_target_msp430_region_lower): New.
* gcc.target/msp430/object-attributes-430.c: New test.
* gcc.target/msp430/object-attributes-default.c: New test.
* gcc.target/msp430/object-attributes-mlarge-any-region.c: New test.
* gcc.target/msp430/object-attributes-mlarge.c: New test.

From-SVN: r276665

* gcc.target/i386/pr71801.c (uuidcache_init): Fix up size of d array.

From-SVN: r276664

* gcc.target/i386/pr71801.c (uuidcache_init): Fix up size of d array.

From-SVN: r276663

call.c (resolve_args): Use cp_expr_loc_or_input_loc in one place.

/cp
2019-10-07  Paolo Carlini  <paolo.carlini@oracle.com>

* call.c (resolve_args): Use cp_expr_loc_or_input_loc in one place.
* decl.c (grokdeclarator): Use id_loc in one place.
* decl2.c (build_anon_union_vars): Use DECL_SOURCE_LOCATION.
* parser.c (cp_parser_delete_expression): Fix the location of the
returned expression.
(cp_parser_throw_expression): Likewise.
* pt.c (determine_specialization): Use DECL_SOURCE_LOCATION.

/testsuite
2019-10-07  Paolo Carlini  <paolo.carlini@oracle.com>

* g++.dg/diagnostic/not-a-function-template-1.C: New.
* g++.dg/template/crash107.C: Adjust expected location.
* g++.dg/template/dependent-expr1.C: Check locations.
* g++.dg/template/error17.C: Check location.

From-SVN: r276661

Introduce can_vcond_compare_p function

z13 supports only non-signaling vector comparisons.  This means we
cannot vectorize LT, LE, GT, GE and LTGT when compiling for z13.
However, we cannot express this restriction today: the code only checks
whether vcond$a$b optab exists, but this does not say anything about the
operation.

Introduce a function that checks whether back-end supports vector
comparisons with individual rtx codes by matching vcond expander's third
argument with a fake comparison with the corresponding rtx code.

gcc/ChangeLog:

2019-10-07  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/77918
* optabs-tree.c (vcond_icode_p): New function.
(vcond_eq_icode_p): Likewise.
(expand_vec_cond_expr_p): Use vcond_icode_p and
vcond_eq_icode_p.
* optabs.c (can_vcond_compare_p): New function.
* optabs.h (can_vcond_compare_p): Likewise.

From-SVN: r276660

Allow COND_EXPR and VEC_COND_EXPR condtions to trap

Right now gimplifier does not allow VEC_COND_EXPR's condition to trap
and introduces a temporary if this could happen, for example, generating

  _5 = _4 > { 2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0 };
  _6 = VEC_COND_EXPR <_5, { -1, -1, -1, -1 }, { 0, 0, 0, 0 }>;

from GENERIC

  VEC_COND_EXPR < (*b > { 2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0 }) ,
                  { -1, -1, -1, -1 } ,
  { 0, 0, 0, 0 } >

This is not necessary and makes the resulting GIMPLE harder to analyze.
Change the gimplifier so as to allow COND_EXPR and VEC_COND_EXPR
conditions to trap.

This patch takes special care to avoid introducing trapping comparisons
in GIMPLE_COND.  They are not allowed, because they would require 3
outgoing edges (then, else and EH), which is awkward to say the least.
Therefore, computations of such conditions should live in their own basic
blocks.

gcc/ChangeLog:

2019-10-07  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/77918
* gimple-expr.c (gimple_cond_get_ops_from_tree): Assert that the
caller passes a non-trapping condition.
(is_gimple_condexpr): Allow trapping conditions.
(is_gimple_condexpr_1): New helper function.
(is_gimple_condexpr_for_cond): New function, acts like old
is_gimple_condexpr.
* gimple-expr.h (is_gimple_condexpr_for_cond): New function.
* gimple.c (gimple_could_trap_p_1): Handle COND_EXPR and
VEC_COND_EXPR. Fix an issue with statements like i = (fp < 1.).
* gimplify.c (gimplify_cond_expr): Use
is_gimple_condexpr_for_cond.
(gimplify_expr): Allow is_gimple_condexpr_for_cond.
* tree-eh.c (operation_could_trap_p): Assert on COND_EXPR and
VEC_COND_EXPR.
(tree_could_trap_p): Handle COND_EXPR and VEC_COND_EXPR.
* tree-ssa-forwprop.c (forward_propagate_into_gimple_cond): Use
is_gimple_condexpr_for_cond, remove pointless tmp check
(forward_propagate_into_cond): Remove pointless tmp check.

From-SVN: r276659

Make gsi_next_nonvirtual_phi do what one expects

gcc/

        * gimple-iterator.h (gsi_next_nonvirtual_phi): Change the semantics to
        match that of other gsi_next_* functions.  Adjust the comment.
        (gsi_start_nonvirtual_phis): New function.
        * ipa-icf.c (sem_function::compare_phi_node): Update uses of
        gsi_next_nonvirtual_phi accordingly.  (No functional change.)

From-SVN: r276658

[doc] -Wuninitialized doesn't do -Wclobbered's job

        * gcc/doc/invoke.texi (-Wuninitialized): Don't mention the clobbered by
        setjmp situation here.  Fix a verb's ending: "the exact variables or
        elements for which there are warnings depends" -> "... depend".

From-SVN: r276657

* ipa-prop.c (ipa_vr::nonzero_p): Add TYPE_UNSIGNED check.

From-SVN: r276655

Disentangle range_fold_*ary_expr() into various independent pieces.

From-SVN: r276654

Use value_range_base::num_pairs instead of vrp_val_is* to check if a range
has one sub-range.

From-SVN: r276653

Delete auto-in_equiv.f90 forgot to use svn delete the first time.

From-SVN: r276651

[IRA] Handle fully-tied destinations in a similar way to earlyclobbers

IRA's make_early_clobber_and_input_conflicts checks for cases in
which an output operand is likely to be an earlyclobber and an input
operand is unlikely to be tieable with it.  If so, the allocno for
the output conflicts with the allocno for the input.  This seems
to work well.

However, a similar situation arises if an output operand is likely
to be tied to one of a set of input operands X and if another input
operand has a different value from all of the operands in X.
E.g. if we have:

  0: "=r, r"
  1: "0, r"
  2: "r, 0"
  3: "r, r"

operand 0 will always be tied to operand 1 or operand 2, so if operand 3
is different from them both, operand 0 acts like an earlyclobber as far
as operand 3 (only) is concerned.  The same is true for operand 2 in:

  0: "=r"
  1: "0"
  2: "r"

In the second example, we'd normally have a copy between operand 1 and
operand 0 if operand 1 dies in the instruction, and so there's rarely
a problem.  But if operand 1 doesn't die in the instruction, operand 0
still acts as an earlyclobber for operand 2 (if different from operand 1),
since in that case LRA must copy operand 1 to operand 0 before the
instruction.

As the existing comment says:

    Avoid introducing unnecessary conflicts by checking classes of the
    constraints and pseudos because otherwise significant code
    degradation is possible for some targets.

I think that's doubly true here.  E.g. it's perfectly reasonable to have
constraints like:

  0: "=r, r"
  1: "0, r"
  2: "r, r"

on targets like s390 that have shorter instructions for tied operands,
but that don't want the size difference to influence RA too much.
We shouldn't treat operand 0 as earlyclobber wrt operand 2 in that case.

This patch therefore treats a normal tied non-earlyclobber output as
being effectively earlyclobber wrt to an input if it is so for *all*
preferred alternatives.

2019-10-07  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* ira-lives.c (check_and_make_def_conflict): Handle cases in which
DEF is not a true earlyclobber but is tied to a specific input
operand, and so is effectively earlyclobber wrt inputs that have
different values.
(make_early_clobber_and_input_conflicts): Pass this case to the above.

From-SVN: r276650

Use CONSTEXPR in machmode.h

This patch adds CONSTEXPR to constructors and operators in machmode.h,
excluding the controversial constexpr-resistant "() {}" constructors,
which would become "= default" if we required C++11.

Some of the other routines could potentially be constexpr too,
but that can be done on an as-needed basis.

2019-10-07  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* machmode.h (opt_mode): Mark constructors with CONSTEXPR.
(pod_mode): Mark operators likewise.
(scalar_int_mode): Mark non-default constructors and
operators with CONSTEXPR.
(scalar_float_mode, scalar_mode, complex_mode): Likewise.
(fixed_size_mode): Likewise.

From-SVN: r276649

[i386] Make the vzeroupper pattern describe its effects (PR91994)

The problem in this PR was that vzeroupper has an effect on register
contents, but those effects weren't modelled in the rtl pattern,
which was just an unspec_volatile.

This patch fixes that by running a subpass after vzeroupper insertion
to add SETs and CLOBBERs as appropriate.  See the comments in the patch
for more details.

2019-10-07  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR target/91994
* config/i386/sse.md (avx_vzeroupper): Turn into a define_expand
and wrap the unspec_volatile in a parallel.
(*avx_vzeroupper): New define_insn.  Use a match_parallel around
the unspec_volatile.
* config/i386/predicates.md (vzeroupper_pattern): Expect the
unspec_volatile to be wrapped in a parallel.
* config/i386/i386-features.c (ix86_add_reg_usage_to_vzeroupper)
(ix86_add_reg_usage_to_vzerouppers): New functions.
(rest_of_handle_insert_vzeroupper): Use them to add register
usage information to the vzeroupper instructions.

gcc/testsuite/
PR target/91994
* gcc.target/i386/pr91994.c: New test.

From-SVN: r276648

re PR tree-optimization/91975 (worse code for small array copy using pointer arithmetic than array indexing)

2019-10-07  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91975
* tree-ssa-loop-ivcanon.c (constant_after_peeling): Consistently
handle invariants.

* g++.dg/tree-ssa/ivopts-3.C: Adjust.
* gcc.dg/vect/vect-profile-1.c: Disable cunrolli.
* gcc.dg/vect/vect-double-reduc-6.c: Disable unrolling of
the innermost loop.
* gcc.dg/vect/vect-93.c: Likewise.
* gcc.dg/vect/vect-105.c: Likewise.
* gcc.dg/vect/pr79920.c: Likewise.
* gcc.dg/vect/no-vfa-vect-102.c: Likewise.
* gcc.dg/vect/no-vfa-vect-101.c: Likewise.
* gcc.dg/vect/pr83202-1.c: Operate on a larger array.
* gfortran.dg/vect/vect-8.f90: Likewise.
* gcc.dg/tree-ssa/cunroll-2.c: Scan early unrolling dump instead
of late one.

From-SVN: r276645

Daily bump.

From-SVN: r276644

Add std::copy_n istreambuf_iterator specialization

Commit this missing part which had been rejected.

From-SVN: r276640

Fix EXECUTE_IF_SET_IN_HARD_REG_SET use

As Maxim pointed out, r276309 broke -O3 bootstrap for aarch64-linux-gnu.
It turns out that EXECUTE_IF_SET_IN_HARD_REG_SET isn't safe to use for
temporary objects because their lifetimes don't extend to the end of
the loop.  This patch therefore uses an explicit temporary instead.

Other patches in the series already used this style (but for
full_reg_and_partial_clobbers rather than full_reg_clobbers),
so AFAICT this is the only place that needs fixing.

2019-10-06  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* var-tracking.c (dataflow_set_clear_at_call): Hoist temporary
function result outside of EXECUTE_IF_SET_IN_HARD_REG_SET.

From-SVN: r276639

Add std::copy_n istreambuf_iterator specialization

* include/bits/stl_algo.h
(__copy_n_a(_IIte, _Size, _OIte)): New.
(__copy_n_a(istreambuf_iterator<>, _Size, _CharT*)): New declaration.
(__copy_n(_IIte, _Size, _OIte, input_iterator_tag)): Adapt to use
latter.
* include/bits/streambuf_iterator.h (istreambuf_iterator<>): Declare
std::__copy_n_a friend.
(__copy_n_a(istreambuf_iterator<>, _Size, _CharT*)): New.
* testsuite/25_algorithms/copy_n/istreambuf_iterator/1.cc: New.
* testsuite/25_algorithms/copy_n/istreambuf_iterator/1_neg.cc: New.
* testsuite/25_algorithms/copy_n/istreambuf_iterator/2_neg.cc: New.

From-SVN: r276638

Add C++11 __iterator_category_t template alias.

* include/bits/stl_iterator_base_types.h (__iterator_category_t): Define
for C++11.
(_RequireInputIte): Likewise and use __enable_if_t.
* include/std/numeric
(__is_random_access_iter): Use __iterator_category_t.

From-SVN: r276637

Add std::copy_n __glibcxx_requires_can_increment checks.

* include/bits/stl_algo.h (copy_n): Add __glibcxx_requires_can_increment
debug checks.
* testsuite/25_algorithms/copy_n/debug/1_neg.cc: New.
* testsuite/25_algorithms/copy_n/debug/2_neg.cc: New.

From-SVN: r276636

[Darwin] Fix some format-related build warnings.

This fixes two error messages to avoid punctuation and contracted
negations.

gcc/ChangeLog:

2019-10-06  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (darwin_override_options): Adjust objective-c
ABI version error messages to avoid punctuation and contracted
negations.

From-SVN: r276635

Daily bump.

From-SVN: r276634

2019-0105  Steven G. Kargl  <kargl@gcc.gnu.org>

In my previous commit, I managed to include a duplicate copy
of the ChangeLog file.  This commit removes the duplicate
portion.

From-SVN: r276630

ipa-inline.c: Fix type; compute size rather than self_size for size of caller function.

* ipa-inline.c: Fix type; compute size rather than self_size
for size of caller function.

From-SVN: r276629

re PR fortran/47054 (Compilation error when cray pointers are declared in both host and internal subroutines)

2019-10-05  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/47045
* decl.c (variable_decl): Do not search parent namespace for symbol.

2019-10-05  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/47045
* gfortran.dg/pr47054_1.f90: New test
* gfortran.dg/pr47054_2.f90: Ditto.

From-SVN: r276627

re PR fortran/91497 (-Wconversion warns when doing explicit type conversion)

2019-10-05  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91497
* gfortran.dg/pr91497.f90: Run on i?86-*-* and x86_64-*-* only.

From-SVN: r276626

re PR fortran/91926 (assumed rank optional)

2019-10-05  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/91926
* trans-expr.c (gfc_conv_gfc_desc_to_cfi_desc): Correct the
assignment of the attribute field to account correctly for an
assumed shape dummy. Assign separately to the gfc and cfi
descriptors since the atribute can be different. Add btanch to
correctly handle missing optional dummies.

2019-10-05  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/91926
* gfortran.dg/ISO_Fortran_binding_13.f90 : New test.
* gfortran.dg/ISO_Fortran_binding_13.c : Additional source.
* gfortran.dg/ISO_Fortran_binding_14.f90 : New test.

2019-10-05  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/91926
* runtime/ISO_Fortran_binding.c (cfi_desc_to_gfc_desc): Do not
modify the bounds and offset for CFI_other.

From-SVN: r276624

[Darwin] Amend section for constants with reloactions.

Darwin's linker doesn't like text section relocations (they require special
enabling).  At least, the Fortran FE seems to generate cases where the
initialiser for a pointer constant can need a relocation.  We can handle
this by special-casing SECCAT_RODATA when the relocation is present by
placing the constant in the .const_data section.

gcc/ChangeLog:

2019-10-05  Iain Sandoe  <iain@sandoe.co.uk>

PR target/59888
* config/darwin.c (darwin_rodata_section): Add relocation flag,
choose const_data section for constants with relocations.
(machopic_select_section): Pass relocation flag to
darwin_rodata_section ().

From-SVN: r276623

PR c++/91369 - Implement P0784R7: constexpr new

PR c++/91369 - Implement P0784R7: constexpr new
c-family/
* c-cppbuiltin.c (c_cpp_builtins): Predefine
__cpp_constexpr_dynamic_alloc=201907 for -std=c++2a.
cp/
* cp-tree.h (enum cp_tree_index): Add CPTI_HEAP_UNINIT_IDENTIFIER,
CPTI_HEAP_IDENTIFIER and CPTI_HEAP_DELETED_IDENTIFIER.
(heap_uninit_identifier, heap_identifier, heap_deleted_identifier):
Define.
(type_has_constexpr_destructor, build_new_constexpr_heap_type,
cxx_constant_dtor): Declare.
* class.c (type_maybe_constexpr_default_constructor): Make static.
(type_maybe_constexpr_destructor, type_has_constexpr_destructor): New
functions.
(finalize_literal_type_property): For c++2a, don't clear
CLASSTYPE_LITERAL_P for types without trivial destructors unless they
have non-constexpr destructors.
(explain_non_literal_class): For c++2a, complain about non-constexpr
destructors rather than about non-trivial destructors.
* constexpr.c: Include stor-layout.h.
(struct constexpr_global_ctx): New type.
(struct constexpr_ctx): Add global field, remove values and
constexpr_ops_count.
(cxx_replaceable_global_alloc_fn): New inline function.
(cxx_eval_call_expression): For c++2a allow calls to replaceable
global allocation functions, for new return address of a heap uninit
var, for delete record its deletion.  Change ctx->values->{get,put} to
ctx->global->values.{get,put}.
(non_const_var_error): Add auto_diagnostic_group sentinel.  Emit
special diagnostics for heap variables.
(cxx_eval_store_expression): Change ctx->values->{get,put} to
ctx->global->values.{get,put}.
(cxx_eval_loop_expr): Initialize jump_target if NULL.  Change
new_ctx.values->remove to ctx->global->values.remove.
(cxx_eval_constant_expression): Change *ctx->constexpr_ops_count
to ctx->global->constexpr_ops_count.  Change ctx->values->{get,put} to
ctx->global->values.{get,put}.
<case NOP_EXPR>: Formatting fix.  On cast of replaceable global
allocation function to some pointer type, adjust the type of
the heap variable and change name from heap_uninit_identifier
to heap_identifier.
(find_heap_var_refs): New function.
(cxx_eval_outermost_constant_expr): Add constexpr_dtor argument,
handle evaluation of constexpr dtors and add tracking of heap
variables.  Use tf_no_cleanup for get_target_expr_with_sfinae.
(cxx_constant_value): Adjust cxx_eval_outermost_constant_expr caller.
(cxx_constant_dtor): New function.
(maybe_constant_value, fold_non_dependent_expr_template,
maybe_constant_init_1): Adjust cxx_eval_outermost_constant_expr
callers.
(potential_constant_expression_1): Ignore clobbers.  Allow
COND_EXPR_IS_VEC_DELETE for c++2a.
* decl.c (initialize_predefined_identifiers): Add heap identifiers.
(decl_maybe_constant_destruction): New function.
(cp_finish_decl): Don't clear TREE_READONLY for constexpr variables
with non-trivial, but constexpr destructors.
(register_dtor_fn): For constexpr variables with constexpr non-trivial
destructors call cxx_maybe_build_cleanup instead of adding destructor
calls at runtime.
(expand_static_init): For constexpr variables with constexpr
non-trivial destructors call cxx_maybe_build_cleanup.
(grokdeclarator): Allow constexpr destructors for c++2a.  Formatting
fix.
(cxx_maybe_build_cleanup): For constexpr variables with constexpr
non-trivial destructors call cxx_constant_dtor instead of adding
destructor calls at runtime.
* init.c: Include stor-layout.h.
(build_new_constexpr_heap_type, maybe_wrap_new_for_constexpr): New
functions.
(build_new_1): For c++2a and new[], add cast around the alloc call
to help constexpr evaluation figure out the type of the heap storage.
(build_vec_delete_1): Set DECL_INITIAL of tbase and emit a DECL_EXPR
for it instead of initializing an uninitialized variable.
* method.c: Include intl.h.
(SFK_CTOR_P, SFK_DTOR_P, SFK_ASSIGN_P, SFK_COPY_P, SFK_MOVE_P): Move
definitions earlier.
(process_subob_fn): Add sfk argument, adjust non-constexpr call
diagnostics based on it.
(walk_field_subobs): Formatting fixes.  Adjust process_subob_fn caller.
(synthesized_method_base_walk): Likewise.
(synthesized_method_walk): Set *constexpr_p to true for dtors in c++2a.
Fix up DR number in comment.
(implicitly_declare_fn): Formatting fix.
* typeck2.c (store_init_value): Don't call cp_fully_fold_init on
initializers of automatic non-constexpr variables in constexpr
functions.
testsuite/
* g++.dg/cpp0x/constexpr-delete2.C: Adjust expected diagnostics for
c++2a.
* g++.dg/cpp0x/locations1.C: Only expect constexpr ~S() diagnostics
in c++17_down, adjust expected wording.
* g++.dg/cpp1y/constexpr-new.C: Only expect diagnostics in c++17_down.
* g++.dg/cpp2a/constexpr-dtor1.C: New test.
* g++.dg/cpp2a/constexpr-dtor2.C: New test.
* g++.dg/cpp2a/constexpr-dtor3.C: New test.
* g++.dg/cpp2a/constexpr-new1.C: New test.
* g++.dg/cpp2a/constexpr-new2.C: New test.
* g++.dg/cpp2a/constexpr-new3.C: New test.
* g++.dg/cpp2a/constexpr-new4.C: New test.
* g++.dg/cpp2a/feat-cxx2a.C: Add __cpp_constinit and
__cpp_constexpr_dynamic_alloc tests.  Tweak __cpp_* tests for c++2a
features to use style like older features, including #ifdef test.
* g++.dg/ext/is_literal_type3.C: New test.

From-SVN: r276622

re PR tree-optimization/91734 (gcc skip an if statement  with "-O1 -ffast-math")

PR tree-optimization/91734
* generic-match-head.c: Include fold-const-call.h.
* match.pd (sqrt(x) cmp c): Check the boundary value and
in case inexact computation of c*c affects comparison of the boundary,
turn LT_EXPR into LE_EXPR, GE_EXPR into GT_EXPR, LE_EXPR into LT_EXPR
or GT_EXPR into GE_EXPR.  Punt for sqrt comparisons against NaN and
for -frounding-math.  For c2, try the next smaller or larger floating
point constant depending on comparison code and if it has the same
sqrt as c2, use it instead of c2.

* gcc.dg/pr91734.c: New test.

From-SVN: r276621

Daily bump.

From-SVN: r276618

Add strftime format checking support for C2x %OB and %Ob (bug 82752).

C2x adds strftime %OB and %Ob formats, for alternative forms of month
names (for mainly Slavic languages where a month name on its own is
declined differently from a month name together with a date within
that month).  This patch adds corresponding format checking support.
(glibc support for these formats was added in glibc 2.27.)

Bootstrapped with no regressions on x86_64-pc-linux-gnu.

PR c/82752
gcc/c-family:
* c-format.c (C_STD_VER): Handle C2x.
(C_STD_NAME): Likewise.
(strftime_flag_specs): Add 'O' modifier with 'p' flag.
(time_char_table): Use separate entry for 'B' and 'b', with 'O'
modifier allowed and 'p' flag.
* c-format.h (enum format_std_version): Add STD_C2X.
(struct format_char_info): Mention 'p' in comment on flags2.

gcc/testsuite:
* gcc.dg/format/c2x-strftime-1.c: New test.

From-SVN: r276605

PR middle-end/91977 - missing -Wstringop-overflow on memcpy into a pointer plus offset

gcc/ChangeLog:

PR middle-end/91977
* tree-ssa-strlen.c (count_nonzero_bytes): Handle assignments with
MEM_REF right operand.  Avoid failing for MEM_REF assignments from
uninitialized objects.

gcc/testsuite/ChangeLog:

PR middle-end/91977
* gcc.dg/Wstringop-overflow-18.c: New test.

From-SVN: r276603

builtins.c (compute_objsize): Add an argument.

gcc/ChangeLog:

* builtins.c (compute_objsize): Add an argument.
* tree-object-size.c (addr_object_size): Same.
(compute_builtin_object_size): Same.
* tree-object-size.h (compute_builtin_object): Same.

gcc/testsuite/ChangeLog:

* gcc.dg/Wstringop-overflow-17.c: New test.

From-SVN: r276602

PR fortran.91959

2019-10-04  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran.91959
* fortran/decl.c (variable_decl): Re-arrange code for matching %FILL.

2019-10-04  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran.91959
* gfortran.dg/pr91959.f90: New test.

From-SVN: r276601

Improve _GLIBCXX_DEBUG safe iterator range size computation.

* include/debug/forward_list
(_Sequence_traits<__debug::forward_list<>>::_S_size): Returns __dp_sign
distance when not empty.
* include/debug/list (_Sequence_traits<__debug::list<>>::_S_size):
Likewise.
* include/debug/helper_functions.h (__dp_sign_max_size): New
_Distance_precision enum entry.
(__valid_range_aux(_IIte, _IIte, _Distance_traits<>::__type,
__false_type)): Adapt.
* include/debug/safe_iterator.tcc
(_Safe_iterator<>::_M_get_distance_to(const _Safe_iterator&)): Review
distance computation.

From-SVN: r276600