Character typenames in errors and warnings

Character type names now incorporate length, kind is only shown if
the default character is not being used.

Examples:

  character(7) is reported as CHARACTER(7)
  character(len=20,kind=4) is reported as CHARACTER(20,4)

dummy character variables with assumed length:

  character(*) is reported as CHARACTER(*)
  character(*,kind=4) is reported as CHARACTER(*,4)

From-SVN: r276505

Makefile.in (OBJS): Add range.o and range-op.o.

* Makefile.in (OBJS): Add range.o and range-op.o.
Remove wide-int-range.o.
* function-tests.c (test_ranges): New.
(function_tests_c_tests): Call test_ranges.
* ipa-cp.c (ipa_vr_operation_and_type_effects): Call
range_fold_unary_expr instead of extract_range_from_unary_expr.
* ipa-prop.c (ipa_compute_jump_functions_for_edge): Same.
* range-op.cc: New file.
* range-op.h: New file.
* range.cc: New file.
* range.h: New file.
* selftest.h (range_tests): New prototype.
* ssa.h: Include range.h.
* tree-vrp.c (value_range_base::value_range_base): New
constructors.
(value_range_base::singleton_p): Do not call
ranges_from_anti_range until sure we will need to.
(value_range_base::type): Rename gcc_assert to
gcc_checking_assert.
(vrp_val_is_max): New argument.
(vrp_val_is_min): Same.
(wide_int_range_set_zero_nonzero_bits): Move from
wide-int-range.cc.
(extract_range_into_wide_ints): Remove.
(extract_range_from_multiplicative_op): Remove.
(extract_range_from_pointer_plus_expr): Abstract POINTER_PLUS code
from extract_range_from_binary_expr.
(extract_range_from_plus_minus_expr): Abstract PLUS/MINUS code
from extract_range_from_binary_expr.
(extract_range_from_binary_expr): Remove.
(normalize_for_range_ops): New.
(range_fold_binary_expr): New.
(range_fold_unary_expr): New.
(value_range_base::num_pairs): New.
(value_range_base::lower_bound): New.
(value_range_base::upper_bound): New.
(value_range_base::upper_bound): New.
(value_range_base::contains_p): New.
(value_range_base::invert): New.
(value_range_base::union_): New.
(value_range_base::intersect): New.
(range_compatible_p): New.
(value_range_base::operator==): New.
(determine_value_range_1): Call range_fold_*expr instead of
extract_range_from_*expr.
* tree-vrp.h (class value_range_base): Add new constructors.
Add methods for union_, intersect, operator==, contains_p,
num_pairs, lower_bound, upper_bound, invert.
(vrp_val_is_min): Add handle_pointers argument.
(vrp_val_is_max): Same.
(extract_range_from_unary_expr): Remove.
(extract_range_from_binary_expr): Remove.
(range_fold_unary_expr): New.
(range_fold_binary_expr): New.
* vr-values.c (vr_values::extract_range_from_binary_expr): Call
range_fold_binary_expr instead of extract_range_from_binary_expr.
(vr_values::extract_range_basic): Same.
(vr_values::extract_range_from_unary_expr): Call
range_fold_unary_expr instead of extract_range_from_unary_expr.
* wide-int-range.cc: Remove.
* wide-int-range.h: Remove.

From-SVN: r276504

Undo unintended change to predicates.md in subversion id 276498

From-SVN: r276503

Daily bump.

From-SVN: r276502

Use the SIGNED_16BIT_OFFSET_EXTRA_P macro for 16-bit signed tests.

2019-10-02   Michael Meissner  <meissner@linux.ibm.com>

* config/rs6000/rs6000.c (mem_operand_gpr): Use
SIGNED_16BIT_OFFSET_EXTRA_P.
(mem_operand_ds_form): Use SIGNED_16BIT_OFFSET_EXTRA_P.
(rs6000_mode_dependent_address): Use SIGNED_16BIT_OFFSET_EXTRA_P.

From-SVN: r276498

Define WIDTH macros for C2x.

As part of the integration of TS 18661-1 into C2x, many features
became unconditional features not depending on any feature test macro
being defined.  This patch updates the conditionals on the *_WIDTH
macros in limits.h and stdint.h accordingly so that they are defined
for C2x.  The macro CR_DECIMAL_DIG in float.h does still require
__STDC_WANT_IEC_60559_BFP_EXT__ to be defined, and a test for this is
added.

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

gcc:
* ginclude/stdint-gcc.h [__STDC_WANT_IEC_60559_BFP_EXT__]: Change
condition on WIDTH macros to [__STDC_WANT_IEC_60559_BFP_EXT__ ||
(__STDC_VERSION__ && __STDC_VERSION__ > 201710L)].
* glimits.h: Likewise.

gcc/testsuite:
* gcc.dg/cr-decimal-dig-2.c: New test.
* gcc.dg/limits-width-2.c: New test.  Based on limits-width-1.c.
* gcc.dg/stdint-width-2.c: New test.  Based on stdint-width-1.c.

From-SVN: r276497

init.c (build_new): Use cp_expr_loc_or_input_loc in two places.

/cp
2019-10-03  Paolo Carlini  <paolo.carlini@oracle.com>

* init.c (build_new): Use cp_expr_loc_or_input_loc in two places.
* name-lookup.c (do_pushdecl): Use DECL_SOURCE_LOCATION.
(push_class_level_binding_1): Likewise.
(set_decl_namespace): Likewise.

/testsuite
2019-10-03  Paolo Carlini  <paolo.carlini@oracle.com>

* g++.dg/diagnostic/integral-array-size-1.C: New.
* g++.dg/cpp0x/alias-decl-1.C: Test location(s) too.
* g++.dg/init/new43.C: Likewise.
* g++.dg/lookup/friend12.C: Likewise.
* g++.dg/lookup/pr79766.C: Likewise.
* g++.dg/lookup/pr84375.C: Likewise.
* g++.dg/other/new-size-type.C: Likewise.

From-SVN: r276496

re PR rtl-optimization/91976 (RTL check: expected code 'const_int', have 'reg' in emit_block_move_hints, at expr.c:1627)

PR rtl-optimization/91976
* expr.c (emit_block_move_hints): Don't call can_move_by_pieces if
size is not CONST_INT_P, set pieces_ok to false in that case.  Simplify
CONST_INT_P (size) && pieces_ok to pieces_ok.  Formatting fix.

From-SVN: r276495

constexpr.c (cxx_eval_constant_expression): If not skipping upon entry to body...

* constexpr.c (cxx_eval_constant_expression) <case CLEANUP_STMT>: If
not skipping upon entry to body, run cleanup with the same *jump_target
as it started to run the cleanup even if the body returns, breaks or
continues.
(potential_constant_expression_1): Allow CLEANUP_STMT.

* g++.dg/ext/constexpr-attr-cleanup1.C: New test.

From-SVN: r276494

constexpr.c (cxx_eval_store_expression): Formatting fix.

* constexpr.c (cxx_eval_store_expression): Formatting fix.  Handle
const_object_being_modified with array type.

From-SVN: r276493

Add changelog entries for r276491.

From-SVN: r276492

PR tree-optimization/80936 - bcmp, bcopy, and bzero not declared nonnull

gcc/testsuite/ChangeLog:

PR tree-optimization/80936
* gcc.dg/Wnonnull-2.c: New test.
* gcc.dg/Wnonnull-3.c: New test.
* gcc.dg/nonnull-3.c: Expect more warnings.

gcc/ChangeLog:

PR tree-optimization/80936
* builtins.def (bcmp, bcopy, bzero): Declare nonnull.

From-SVN: r276491

Fix ALL_REGS thinko in initialisation of function_used_regs

My change to the -fipa-ra bookkeeping used ALL_REGS as the supposedly
safe default assumption, but ALL_REGS isn't literally all registers,
just a close approximation.

This caused a bootstrap failure on arm-linux-gnu, where the condition
code register isn't in ALL_REGS and so was being masked out of some
call-clobbered sets.

2019-10-02  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* cgraph.c (cgraph_node::rtl_info): Use SET_HARD_REG_SET
instead of reg_class_contents[ALL_REGS].

From-SVN: r276489

runtime: mark go-context.S as no-executable-stack and split-stack supported

    The .note.GNU-stack section tells the linker that this object does not
    require an executable stack.

    The .note.GNU-split-stack section tells the linker that functions in
    this object can be called directly by split-stack functions, without
    require a large stack.

    The .note.GNU-no-split-stack section tells the linker that functions
    in this object do not have a split-stack prologue.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/198440

From-SVN: r276488

Improve C++ fold caching efficiency.

While looking at concepts caching I noticed that we were clearing the caches
unnecessarily for non-constant initialization, which shouldn't affect
folding.

* typeck2.c (store_init_value): Only clear_cv_and_fold_caches if the
value is constant.

From-SVN: r276487

Add some hash_map_safe_* functions like vec_safe_*.

gcc/
* hash-map.h (default_hash_map_size): New variable.
(create_ggc): Use it as default argument.
(hash_map_maybe_create, hash_map_safe_get)
(hash_map_safe_get_or_insert, hash_map_safe_put): New fns.
gcc/cp/
* constexpr.c (maybe_initialize_fundef_copies_table): Remove.
(get_fundef_copy): Use hash_map_safe_get_or_insert.
* cp-objcp-common.c (cp_get_debug_type): Use hash_map_safe_*.
* decl.c (store_decomp_type): Remove.
(cp_finish_decomp): Use hash_map_safe_put.
* init.c (get_nsdmi): Use hash_map_safe_*.
* pt.c (store_defaulted_ttp, lookup_defaulted_ttp): Remove.
(add_defaults_to_ttp): Use hash_map_safe_*.

From-SVN: r276484

Testsuite, remove alloca header

2019-10-02  Andreas Tobler  <andreast@gcc.gnu.org>

* testsuite/libgomp.oacc-c-c++-common/loop-default.h: Remove alloca.h
include. Replace alloca () with __builtin_alloca ().
* testsuite/libgomp.oacc-c-c++-common/loop-dim-default.c: Likewise.

Reviewed-by: Thomas Schwinge <thomas@codesourcery.com>
From-SVN: r276479

re PR fortran/91784 (ICE in gfc_real2complex, at fortran/arith.c:2208)

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91784
* simplify.c (gfc_convert_constant): Simplify expression if the
expres ion type is EXPR_OP.

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91784
* gfortran.dg/pr91784.f90: New test.

From-SVN: r276474

re PR fortran/91785 (ICE in check_assumed_size_reference, at fortran/resolve.c:1601)

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91785
* primary.c (gfc_match_varspec): Ensure an inquiry parameter has
it locus set.

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91785
* gfortran.dg/pr91785.f90: New test.

From-SVN: r276473

re PR fortran/91942 (ICE in match_vtag, at fortran/io.c:1485)

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91942
* io.c (match_vtag): Check for non-NULL result->symtree.
(match_out_tag): Check for invalid constant due to inquiry parameter.
(match_filepos): Instead of a syntax error, go to cleanup to get better
error messages.

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91942
* gfortran.dg/pr91587.f90: Update dg-error regex.
* gfortran.dg/pr91942.f90: New test.

From-SVN: r276472

re PR fortran/91943 (ICE in gfc_conv_constant_to_tree, at fortran/trans-const.c:370)

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91943
* match.c (gfc_match_call): BOZ cannot be an actual argument in
a subroutine reference.
* resolve.c (resolve_function): BOZ cannot be an actual argument in
a function reference.

2019-10-02  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91943
gfortran.dg/pr91943.f90

From-SVN: r276471

cif-code.def (MAX_INLINE_INSNS_SINGLE_O2_LIMIT, [...]): New.

* cif-code.def (MAX_INLINE_INSNS_SINGLE_O2_LIMIT,
MAX_INLINE_INSNS_AUTO_O2_LIMIT): New.
* ipa-inline.c (inline_insns_single, inline_insns_auto): New functions.
(can_inline_edge_by_limits_p): Use it.
(big_speedup_p): Use PARAM_INLINE_MIN_SPEEDUP_O2.
(want_inline_small_function_p): Use O2 bounds.
(edge_badness): LIkewise.
* opts.c (default_options): Add OPT_finline_functions.
* params.def (PARAM_INLINE_MIN_SPEEDUP_O2,
PARAM_MAX_INLINE_INSNS_SINGLE_O2, PARAM_MAX_INLINE_INSNS_AUTO_O2):
New parameters.

* g++.dg/tree-ssa/pr53844.C: Add -fno-inline-functions --param
max-inline-insns-single-O2=200.
* gcc.c-torture/execute/builtins/builtins.exp: Add
-fno-inline-functions to additional_flags.
* gcc.dg/ipa/inline-7.c: Add -fno-inline-functions.
* gcc.dg/optimize-bswapsi-5.c: Add -fno-inline-functions.
* gcc.dg/tree-ssa/ssa-thread-12.c: Add --param
early-inlining-insns-O2=14 -fno-inline-functions; revert previous
change.
* gcc.dg/winline-3.c: Use --param max-inline-insns-single-O2=1
--param inline-min-speedup-O2=100
instead of --param max-inline-insns-single=1 --param
inline-min-speedup=100

* invoke.texi (-finline-functions): Update documentation.
(max-inline-insns-single-O2, max-inline-insns-auto-O2,
inline-min-speedup-O2): Document.
(early-inlining-insns-O2): Simplify docs.

From-SVN: r276470

cif-code.def (MAX_INLINE_INSNS_SINGLE_O2_LIMIT, [...]): New.

* cif-code.def (MAX_INLINE_INSNS_SINGLE_O2_LIMIT,
MAX_INLINE_INSNS_AUTO_O2_LIMIT): New.
* ipa-inline.c (inline_insns_single, inline_insns_auto): New functions.
(can_inline_edge_by_limits_p): Use it.
(big_speedup_p): Use PARAM_INLINE_MIN_SPEEDUP_O2.
(want_inline_small_function_p): Use O2 bounds.
(edge_badness): LIkewise.
* opts.c (default_options): Add OPT_finline_functions.
* params.def (PARAM_INLINE_MIN_SPEEDUP_O2,
PARAM_MAX_INLINE_INSNS_SINGLE_O2, PARAM_MAX_INLINE_INSNS_AUTO_O2):
New parameters.

* g++.dg/tree-ssa/pr53844.C: Add -fno-inline-functions --param
max-inline-insns-single-O2=200.
* gcc.c-torture/execute/builtins/builtins.exp: Add
-fno-inline-functions to additional_flags.
* gcc.dg/ipa/inline-7.c: Add -fno-inline-functions.
* gcc.dg/optimize-bswapsi-5.c: Add -fno-inline-functions.
* gcc.dg/tree-ssa/ssa-thread-12.c: Add --param
early-inlining-insns-O2=14 -fno-inline-functions; revert previous
change.
* gcc.dg/winline-3.c: Use --param max-inline-insns-single-O2=1
--param inline-min-speedup-O2=100
instead of --param max-inline-insns-single=1 --param
inline-min-speedup=100

* invoke.texi (-finline-functions): Update documentation.
(max-inline-insns-single-O2, max-inline-insns-auto-O2,
inline-min-speedup-O2): Document.
(early-inlining-insns-O2): Simplify docs.

From-SVN: r276469

Remove greedy wildcards from libstdc++ linker script

The only symbols matched by std::e[a-q]* and std::e[s-z]* that are
supposed to be in the GLIBCXX_3.4 version are std::exception::* and
std::endl and std::ends. The latter two already have explicit patterns
matching them, so we just need to match std::exception::*.

This change ensures that any new symbols with a return type of
std::enable_if<...> are not added to the GLIBCXX_3.4 version.

* config/abi/pre/gnu.ver: Tighten up greedy wildcards.

From-SVN: r276468

ifcvt: improve cost estimation (PR 87047)

PR rtl-optimization/87047
* ifcvt.c (average_cost): New static function.  Use it...
(noce_process_if_block): ... here.

testsuite/
* gcc.dg/pr87047.c: New test.

From-SVN: r276466

[PR testsuite/91842] Skip gcc.dg/ipa/ipa-sra-19.c on power

2019-10-02  Martin Jambor  <mjambor@suse.cz>

PR testsuite/91842
* gcc.dg/ipa/ipa-sra-19.c: Skip on powerpc.

From-SVN: r276465

module.c (load_commons): Initialize flags to 0 to silecne -Wmaybe-uninitialized warning.

* module.c (load_commons): Initialize flags to 0 to silecne
-Wmaybe-uninitialized warning.
(read_module): Likewise for n and comp_name.

From-SVN: r276464

Document non-conformance of parallel mode to recent C++ standards

* doc/xml/manual/parallel_mode.xml: Add caveat about support for
recent standards.
* doc/html/*: Regenerate.

From-SVN: r276463

rs6000-protos.h (expand_block_move): Change prototype.

2019-10-02  Aaron Sawdey <acsawdey@linux.ibm.com>

* config/rs6000/rs6000-protos.h (expand_block_move): Change prototype.
* config/rs6000/rs6000-string.c (expand_block_move): Add
might_overlap parm.
* config/rs6000/rs6000.md (movmemsi): Add new pattern.
(cpymemsi): Add might_overlap parm to expand_block_move() call.

From-SVN: r276462

builtins.c (expand_builtin_memory_copy_args): Add might_overlap parm.

2019-10-02  Aaron Sawdey <acsawdey@linux.ibm.com>

* builtins.c (expand_builtin_memory_copy_args): Add might_overlap parm.
(expand_builtin_memcpy): Use might_overlap parm.
(expand_builtin_mempcpy_args): Use might_overlap parm.
(expand_builtin_memmove): Call expand_builtin_memory_copy_args.
(expand_builtin_memory_copy_args): Add might_overlap parm.
* expr.c (emit_block_move_via_cpymem): Rename to
emit_block_move_via_pattern, add might_overlap parm, use cpymem
or movmem optab as appropriate.
(emit_block_move_hints): Add might_overlap parm, do the right
thing for might_overlap==true.
* expr.h (emit_block_move_hints): Update prototype.

From-SVN: r276461

tree-eh.h (unsplit_eh_edges): Declare.

* tree-eh.h (unsplit_eh_edges): Declare.
* tree-eh.c (maybe_remove_unreachable_handlers): Detect more cases.
(unsplit_eh_edges): New function wrapping unsplit_all_eh.
* gimple-ssa-store-merging.c: Include cfganal.h cfgcleanup.h except.h.
(struct store_immediate_info): Add lp_nr field.
(store_immediate_info::store_immediate_info): Add NR2 parameter and
initialize lp_nr with it.
(struct merged_store_group): Add lp_nr and only_constants fields.
(merged_store_group::merged_store_group): Initialize them.
(merged_store_group::can_be_merged_into): Deal with them.
(pass_store_merging): Rename terminate_and_release_chain into
terminate_and_process_chain.
(pass_store_merging::terminate_and_process_all_chains): Adjust to above
renaming and remove useless assertions.
(pass_store_merging::terminate_all_aliasing_chains): Small tweak.
(stmts_may_clobber_ref_p): Be prepared for different basic blocks.
(imm_store_chain_info::coalesce_immediate_stores): Use only_constants
instead of always recomputing it and compare lp_nr.
(imm_store_chain_info::output_merged_store): If the group is in an
active EH region, register new stores if they can throw.  Moreover,
if the insertion has created new basic blocks, adjust the PHI nodes
of the post landing pad.
(imm_store_chain_info::output_merged_stores): If the original stores
are in an active EH region, deregister them.
(lhs_valid_for_store_merging_p): Prettify.
(adjust_bit_pos): New function extracted from...
(mem_valid_for_store_merging): ...here.  Use it for the base address
and also for the offset if it is the addition of a constant.
(lp_nr_for_store): New function.
(pass_store_merging::process_store): Change return type to bool.
Call lp_nr_for_store to initialize the store info.  Propagate the
return status of various called functions to the return value.
(store_valid_for_store_merging_p): New predicate.
(enum basic_block_status): New enumeration.
(get_status_for_store_merging): New function.
(pass_store_merging::execute): If the function can throw and catch
non-call exceptions, unsplit the EH edges on entry and clean up the
CFG on exit if something changed.  Call get_status_for_store_merging
for every basic block and keep the chains open across basic blocks
when possible.  Terminate and process open chains at the end, if any.

From-SVN: r276459

Fix shadowing in globalize_reg

2019-10-02  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* reginfo.c (globalize_reg): Fix shadowed variable in
function_abis walk.

From-SVN: r276457

Fix MIPS call-clobbered-*.c tests

Jeff pointed out that gcc.target/mips/call-clobbered-4.c started
failing after the function-abi series.  This is because IRA used
to treat partly call-clobbered registers as hard conflicts and
so wouldn't consider them for -fcaller-saves.  Now that we treat
call clobbers the same way regardless of where they come from,
we can use $f21 as a caller-save register.  This in turn means
that -Os is no longer a special case in call-clobbered-3.c.

(The new code is the same size as the old code.)

2019-10-02  Richard Sandiford  <richard.sandiford@arm.com>

gcc/testsuite/
* gcc.target/mips/call-clobbered-3.c: Remove skip for -Os.
* gcc.target/mips/call-clobbered-4.c: Delete.

From-SVN: r276456

[PATCH] Do not check call type compatibility when cloning cgraph-edges

2019-10-02  Martin Jambor  <mjambor@suse.cz>

* cgraph.c (symbol_table::create_edge): New parameter cloning_p,
do not compute some stuff when set.
(cgraph_node::create_edge): Likewise.
(cgraph_node::create_indirect_edge): Renamed last parameter to
coning_p and flipped its meaning, don't even calculate
inline_failed when set.
* cgraph.h (cgraph_node::create_edge): Add new parameter.
(symbol_table::::create_edge): Likewise.
(cgraph_node::create_indirect_edge): Rename last parameter, flip
the default value.
* cgraphclones.c (cgraph_edge::clone): Pass true cloning_p to all
call graph edge creating functions.

From-SVN: r276455

re PR c++/91222 (507.cactuBSSN_r build fails in warn_types_mismatch at ipa-devirt.c:1006 since r273571)

PR c++/91222
* ipa-devirt.c (warn_types_mismatch): Fix conditional on anonymous
namespace types.

From-SVN: r276454

[ARC] Pass along "-mcode-density" flag to "as"

This change makes sure that if the driver is invoked with
"-mcode-density" flag, then the assembler will receive it
too.

gcc/
xxxx-xx-xx  Shahab Vahedi  <shahab@synopsys.com>

        * config/arc/arc.h (ASM_SPEC): pass -mcode-density

From-SVN: r276453

tree-vectorizer.h (vect_transform_reduction): Declare.

2019-10-02  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (vect_transform_reduction): Declare.
* tree-vect-stmts.c (vect_transform_stmt): Use it.
* tree-vect-loop.c (vectorizable_reduction): Split out reduction
stmt transform to ...
(vect_transform_reduction): ... this.

From-SVN: r276452

Support OpenMP's use_device_addr in Fortran

        gcc/fortran/
* dump-parse-tree.c (show_omp_clauses): Handle OMP_LIST_USE_DEVICE_ADDR.
* gfortran.h (enum): Add OMP_LIST_USE_DEVICE_ADDR.
* openmp.c (omp_mask1): Likewise.
(gfc_match_omp_clauses): Match 'use_device_addr'.
(OMP_TARGET_DATA_CLAUSES): Add OMP_LIST_USE_DEVICE_ADDR.
(resolve_omp_clauses): Add it; add is_device_ptr checks.

        gcc/testsuite/
* gfortran.dg/gomp/is_device_ptr-1.f90: New.

From-SVN: r276449

re PR c++/91606 (Optimization leads to invalid code)

2019-10-02  Richard Biener  <rguenther@suse.de>

PR c++/91606
* decl.c (build_ptrmemfunc_type): Mark pointer-to-member
fat pointer structure members as DECL_NONADDRESSABLE_P.

* g++.dg/torture/pr91606.C: New testcase.

From-SVN: r276448

Improve OMP/ACC error diagnostic in Fortran

gcc/fortran/
* openmp.c (gfc_match_omp_clauses): Show a clause-parsing
error if none was rised before.
* parse.c (matcha, matcho): If error occurred after
OpenMP/OpenACC directive matched, do not try other directives.

gcc/testsuite/
* gfortran.dg/goacc/asyncwait-1.f95: Handle new error message.
* gfortran.dg/goacc/asyncwait-2.f95: Likewise
* gfortran.dg/goacc/asyncwait-3.f95: Likewise
* gfortran.dg/goacc/asyncwait-4.f95: Likewise
* gfortran.dg/goacc/default-2.f: Likewise
* gfortran.dg/goacc/enter-exit-data.f95: Likewise
* gfortran.dg/goacc/if.f95: Likewise
* gfortran.dg/goacc/list.f95: Likewise
* gfortran.dg/goacc/literal.f95: Likewise
* gfortran.dg/goacc/loop-2-kernels-tile.f: Likewise95
* gfortran.dg/goacc/loop-2-parallel-tile.f95: Likewise
* gfortran.dg/goacc/loop-7.f95: Likewise
* gfortran.dg/goacc/parallel-kernels-cla: Likewiseuses.f95
* gfortran.dg/goacc/routine-6.f90: Likewise
* gfortran.dg/goacc/several-directives.f95: Likewise
* gfortran.dg/goacc/sie.f95: Likewise
* gfortran.dg/goacc/tile-1.f90: Likewise
* gfortran.dg/goacc/update-if_present-2.: Likewisef90
* gfortran.dg/gomp/declare-simd-1.f90: Likewise
* gfortran.dg/gomp/pr29759.f90: Likewise

From-SVN: r276447

Fix coding style comment, missed in last commit

        * trans-openmp.c (gfc_omp_is_optional_argument): Fix coding
        style.

From-SVN: r276446

Fix omp target issue with Fortran optional arguments

        gcc/
        * omp-low.c (lower_omp_target): Dereference optional argument
        to work with the right pointer.

        gcc/testsuite/
        * libgomp/testsuite/libgomp.fortran/use_device_ptr-optional-1.f90: New.

From-SVN: r276445

f95-lang.c (LANG_HOOKS_OMP_IS_OPTIONAL_ARGUMENT): Define to gfc_omp_is_optional_argument.

2019-10-02  Kwok Cheung Yeung  <kcy@codesourcery.com>

        gcc/fortran/
        * f95-lang.c (LANG_HOOKS_OMP_IS_OPTIONAL_ARGUMENT): Define to
        gfc_omp_is_optional_argument.
        * trans-decl.c (create_function_arglist): Set
        GFC_DECL_OPTIONAL_ARGUMENT in the generated decl if the parameter is
        optional.
        * trans-openmp.c (gfc_omp_is_optional_argument): New.
        (gfc_omp_privatize_by_reference): Return true if the decl is an
        optional pass-by-reference argument.
        * trans.h (gfc_omp_is_optional_argument): New declaration.
        (lang_decl): Add new optional_arg field.
        (GFC_DECL_OPTIONAL_ARGUMENT): New macro.

        gcc/
        * langhooks-def.h (LANG_HOOKS_OMP_IS_OPTIONAL_ARGUMENT): Default to
        false.
        (LANG_HOOKS_DECLS): Add LANG_HOOKS_OMP_IS_OPTIONAL_ARGUMENT.
        * langhooks.h (omp_is_optional_argument): New hook.
        * omp-general.c (omp_is_optional_argument): New.
        * omp-general.h (omp_is_optional_argument): New declaration.
        * omp-low.c (lower_omp_target): Create temporary for received value
        and take the address for new_var if the original variable was a
        DECL_BY_REFERENCE.  Use size of referenced object when a
        pass-by-reference optional argument used as argument to firstprivate.

From-SVN: r276444

internal.h (enum include_type): Remove trailing comma.

2019-10-02  Richard Biener  <rguenther@suse.de>

* internal.h (enum include_type): Remove trailing comma.

From-SVN: r276443

re PR tree-optimization/91940 (__builtin_bswap16 loop optimization)

PR tree-optimization/91940
* tree-vect-patterns.c: Include tree-vector-builder.h and
vec-perm-indices.h.
(vect_recog_rotate_pattern): Also handle __builtin_bswap16, either by
unpromoting the argument back to uint16_t, or by converting into a
rotate, or into shifts plus ior.

* gcc.dg/vect/vect-bswap16.c: Add -msse4 on x86, run on all targets,
expect vectorized 1 loops message on both vect_bswap and sse4_runtime
targets.
* gcc.dg/vect/vect-bswap16a.c: New test.

From-SVN: r276442

tree-vectorizer.h (stmt_vec_info_type::cycle_phi_info_type): New.

2019-10-02  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (stmt_vec_info_type::cycle_phi_info_type):
New.
(vect_transform_cycle_phi): Declare.
* tree-vect-stmts.c (vect_transform_stmt): Call
vect_transform_cycle_phi.
* tree-vect-loop.c (vectorizable_reduction): Split out
PHI transformation stage to ...
(vect_transform_cycle_phi): ... here.

From-SVN: r276441

[LRA] Don't make eliminable registers live (PR91957)

One effect of https://gcc.gnu.org/ml/gcc-patches/2019-09/msg00802.html
was to strengthen the sanity check in lra_assigns so that it checks
whether reg_renumber is consistent with the whole conflict set.
This duly tripped on csky for a pseudo that had been allocated
to the eliminated frame pointer.  (csky doesn't have a separate
hard frame pointer.)

lra-lives uses:

/* Set of hard regs (except eliminable ones) currently live.  */
static HARD_REG_SET hard_regs_live;

to track the set of live directly-referenced hard registers, and it
correctly implements the exclusion when setting up the initial set:

  hard_regs_live &= ~eliminable_regset;

But later calls to make_hard_regno_live and make_hard_regno_dead
would process eliminable registers like other registers, recording
conflicts for them and potentially making them live.  (Note that
after r266086, make_hard_regno_dead adds conflicts for registers
that are already marked dead.)  I think this would have had the
effect of pessimising targets without a separate hard frame pointer.

2019-10-02  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR middle-end/91957
* lra-lives.c (make_hard_regno_dead): Don't record conflicts for
eliminable registers.
(make_hard_regno_live): Likewise, and don't make them live.

From-SVN: r276440

re PR libfortran/91593 (Implicit enum conversions in libgfortran/io/transfer.c)

2019-10-01  Jerry DeLisle  <jvdelisle@gcc.gnu.org>

PR libfortran/91593
* io/read.c (read_decimal): Cast constant to size_t to turn off
a bogus warning.
* io/write.c (btoa_big): Use memset in lieu of setting the null
byte in a string buffer to turn off a bogus warning.

From-SVN: r276439

Daily bump.

From-SVN: r276438

Handle :: tokens in C for C2x.

As part of adding [[]]-style attributes, C2x adds the token :: for use
in scoped attribute names.

This patch adds corresponding support for that token in C to GCC.  The
token is supported both for C2x and for older gnu* standards (on the
basis that extensions are normally supported in older gnu* versions;
people will expect to be able to use [[]] attributes, before C2x is
the default, without needing to use -std=gnu2x).

There are no cases in older C standards where the token : can be
followed by a token starting with : in syntactically valid sources;
the only cases the :: token could break in older standard C thus are
ones involving concatenation of pp-tokens where the result does not
end up as tokens (e.g., gets stringized).  In GNU C extensions, the
main case where :: might appear in existing sources is in asm
statements, and the C parser is thus made to handle it like two
consecutive : tokens, which the C++ parser already does.  A limited
test of various positionings of :: in asm statements is added to the
testsuite (in particular, to cover the syntax error when :: means too
many colons but a single : would be OK), but existing tests cover a
variety of styles there anyway.

Technically there are cases in Objective-C and OpenMP for which this
also changes how previously valid code is lexed: the objc-selector-arg
syntax allows multiple consecutive : tokens (although I don't think
they are particularly useful there), while OpenMP syntax includes
array section syntax such as [:] which, before :: was a token, could
also be written as [::> (there might be other OpenMP cases potentially
affected, I didn't check all the OpenMP syntax in detail).  I don't
think either of those cases affects the basis for supporting the ::
token in all -std=gnu* modes, or that there is any obvious need to
special-case handling of CPP_SCOPE tokens for those constructs the way
there is for asm statements.

cpp_avoid_paste, which determines when spaces need adding between
tokens in preprocessed output where there wouldn't otherwise be
whitespace between them (e.g. if stringized), already inserts space
between : and : unconditionally, rather than only for C++, so no
change is needed there (but a C2x test is added that such space is
indeed inserted).

Bootstrapped with no regressions on x86-64-pc-linux-gnu.

gcc/c:
* c-parser.c (c_parser_asm_statement): Handle CPP_SCOPE like two
CPP_COLON tokens.

gcc/testsuite:
* gcc.dg/asm-scope-1.c, gcc.dg/cpp/c11-scope-1.c,
gcc.dg/cpp/c17-scope-1.c, gcc.dg/cpp/c2x-scope-1.c,
gcc.dg/cpp/c2x-scope-2.c, gcc.dg/cpp/c90-scope-1.c,
gcc.dg/cpp/c94-scope-1.c, gcc.dg/cpp/c99-scope-1.c,
gcc.dg/cpp/gnu11-scope-1.c, gcc.dg/cpp/gnu17-scope-1.c,
gcc.dg/cpp/gnu89-scope-1.c, gcc.dg/cpp/gnu99-scope-1.c: New tests.

libcpp:
* include/cpplib.h (struct cpp_options): Add member scope.
* init.c (struct lang_flags, lang_defaults): Likewise.
(cpp_set_lang): Set scope member of pfile.
* lex.c (_cpp_lex_direct): Test CPP_OPTION (pfile, scope) not
CPP_OPTION (pfile, cplusplus) for creating CPP_SCOPE tokens.

From-SVN: r276434

Support prefixes in diagnostic_show_locus

Previously, diagnostic_show_locus saved and restored the pretty_printer's
prefix, clearing it for the duration of the call.

I have a patch kit in development that can benefit from applying a prefix
to the output of d_s_l, so this patch adds support to d_s_l for printing
such prefixes.

It moves the save and restore of the pp's prefix from d_s_l to all of its
callers, and updates diagnostic-show-locus.c to properly handle prefixes.

gcc/c-family/ChangeLog:
* c-opts.c (c_diagnostic_finalizer): Temporarily clear prefix when
calling diagnostic_show_locus, rather than destroying it afterwards.

gcc/ChangeLog:
* diagnostic-show-locus.c (layout::print_gap_in_line_numbering):
Call pp_emit_prefix.
(layout::print_source_line): Likewise.
(layout::start_annotation_line): Likewise.
(diagnostic_show_locus): Remove call to temporarily clear the
prefix.
(selftest::test_one_liner_fixit_remove): Add test coverage for the
interaction of pp_set_prefix with rulers and fix-it hints.
* diagnostic.c (default_diagnostic_finalizer): Temporarily clear
prefix when calling diagnostic_show_locus, rather than destroying
it afterwards.
(print_parseable_fixits): Temporarily clear prefix.
* pretty-print.c (pp_format): Save and restore line_length, rather
than assuming it is zero.
(pp_output_formatted_text): Remove assertion that line_length is
zero.

gcc/fortran/ChangeLog:
* error.c (gfc_diagnostic_starter): Clear the prefix before
calling diagnostic_show_locus.

gcc/testsuite/ChangeLog:
* gcc.dg/plugin/diagnostic_group_plugin.c (test_begin_group_cb):
Clear the prefix before emitting the "END GROUP" line.
* gcc.dg/plugin/diagnostic_plugin_test_show_locus.c
(custom_diagnostic_finalizer): Temporarily clear prefix when
calling diagnostic_show_locus, rather than destroying it
afterwards.

From-SVN: r276433

Make some new algorithms work in parallel mode

* include/experimental/algorithm (experimental::sample): Qualify call
to __sample correctly.
* include/parallel/algo.h (sample, for_each_n): Add using-declarations
for algorithms that don't have parallel implementations.

From-SVN: r276432

Make some parallel mode algorithms usable in constexpr contexts

This makes the __parallel::equal and __parallel:lexicographical_compare
algorithms usable in constant expressions, by dispatching to the
sequential algorithm when calling during constant evaluation.

* include/parallel/algobase.h (equal, lexicographical_compare): Add
_GLIBCXX20_CONSTEXPR and dispatch to sequential algorithm when being
constant evaluated.
* include/parallel/algorithmfwd.h (equal, lexicographical_compare):
Add _GLIBCXX20_CONSTEXPR.

From-SVN: r276431

Disable tests that aren't valid in parallel mode

Tests that depend on debug mode can't be tested in parallel mode.

* testsuite/17_intro/using_namespace_std_tr1_neg.cc: Skip test for
parallel mode.
* testsuite/20_util/hash/84998.cc: Likewise.
* testsuite/23_containers/deque/types/pmr_typedefs_debug.cc: Likewise.
* testsuite/23_containers/forward_list/pmr_typedefs_debug.cc: Likewise.
* testsuite/23_containers/list/pmr_typedefs_debug.cc: Likewise.
* testsuite/23_containers/map/pmr_typedefs_debug.cc: Likewise.
* testsuite/23_containers/multimap/pmr_typedefs_debug.cc: Likewise.
* testsuite/23_containers/multiset/pmr_typedefs_debug.cc: Likewise.
* testsuite/23_containers/set/pmr_typedefs_debug.cc: Likewise.
* testsuite/23_containers/unordered_map/pmr_typedefs_debug.cc:
Likewise.
* testsuite/23_containers/unordered_multimap/pmr_typedefs_debug.cc:
Likewise.
* testsuite/23_containers/unordered_multiset/pmr_typedefs_debug.cc:
Likewise.
* testsuite/23_containers/unordered_set/pmr_typedefs_debug.cc:
Likewise.
* testsuite/23_containers/vector/cons/destructible_debug_neg.cc:
Likewise.
* testsuite/23_containers/vector/types/pmr_typedefs_debug.cc: Likewise.
* testsuite/25_algorithms/binary_search/partitioned.cc: Likewise.
* testsuite/25_algorithms/copy/86658.cc: Likewise.
* testsuite/25_algorithms/equal_range/partitioned.cc: Likewise.
* testsuite/25_algorithms/lexicographical_compare/71545.cc: Likewise.
* testsuite/25_algorithms/lower_bound/partitioned.cc: Likewise.
* testsuite/25_algorithms/upper_bound/partitioned.cc: Likewise.

From-SVN: r276430

Fix non-reserved names in Parallel Mode headers

* include/parallel/algo.h: Replace non-reserved names.
* include/parallel/multiway_merge.h: Likewise.
* include/parallel/multiway_mergesort.h: Likewise.
* include/parallel/numericfwd.h: Likewise.
* testsuite/17_intro/names.cc: Add RAI to test macros.

From-SVN: r276429

Regenerate `liboffloadmic/plugin/configure' for r275564 ("[ARM/FDPIC v6 02/24] [ARM] FDPIC...

Regenerate `liboffloadmic/plugin/configure' for r275564 ("[ARM/FDPIC v6
02/24] [ARM] FDPIC: Handle arm*-*-uclinuxfdpiceabi in configure
scripts") too.

liboffloadmic/
* plugin/configure: Regenerate.

From-SVN: r276428

tree-ssa-alias.c (nonoverlapping_component_refs_since_match_p): Rename to ...

* tree-ssa-alias.c (nonoverlapping_component_refs_since_match_p):
Rename to ...
(nonoverlapping_refs_since_match_p): ... this; handle also
ARRAY_REFs.
(alias_stats): Update stats.
(dump_alias_stats): Likewise.
(cheap_array_ref_low_bound): New function.
(aliasing_matching_component_refs_p): Add partial_overlap
argument;
pass it to nonoverlapping_refs_since_match_p.
(aliasing_component_refs_walk): Update call of
aliasing_matching_component_refs_p
(nonoverlapping_array_refs_p): New function.
(decl_refs_may_alias_p, indirect_ref_may_alias_decl_p,
indirect_refs_may_alias_p): Update calls of
nonoverlapping_refs_since_match_p.

* gcc.dg/tree-ssa/alias-access-path-10.c: New testcase.
* gcc.dg/tree-ssa/alias-access-path-11.c: New testcase.

From-SVN: r276427

re PR target/85401 (segfault building code for VAX)

PR target/85401
* ira-color.c (allocno_copy_cost_saving): Call
ira_init_register_move_cost_if_necessary.

From-SVN: r276426

libada: Respect `--enable-version-specific-runtime-libs'

Respect the `--enable-version-specific-runtime-libs' configuration
option in libada/, so that shared gnatlib libraries will be installed
in non-version-specific $(toolexeclibdir) if requested.  In a
cross-compilation environment this helps setting up a consistent
sysroot, which can then be shared between the host and the target
system.

This lets one have `libgnarl-10.so' and `libgnat-10.so' installed in say
/usr/lib and /usr/$(target_alias)/lib for a native and a cross-build
respectively, rather than in /usr/lib/gcc/$(target_alias)/10.0.0/adalib.

Update the settings of $(toolexecdir) and $(toolexeclibdir), unused till
now, to keep the current arrangement in the version-specific case and
make the new option to be enabled by default, unlike with the other
target libraries, so as to keep existing people's build infrastructure
unaffected.

Of course if someone does use `--disable-version-specific-runtime-libs'
already, then the installation location of shared gnatlib libraries will
change, but presumably this is what they do want anyway as the current
situation where the option is ignored in libada/ only is an anomaly
really rather than one that is expected or desired.

gcc/ada/
* gcc-interface/Makefile.in (ADA_RTL_DSO_DIR): New variable.
(install-gnatlib): Use it in place of ADA_RTL_OBJ_DIR for shared
library installation.

libada/
* Makefile.in (toolexecdir, toolexeclibdir): New variables.
(LIBADA_FLAGS_TO_PASS): Add `toolexeclibdir'.
* configure.ac: Add `--enable-version-specific-runtime-libs'.
Update version-specific `toolexecdir' and `toolexeclibdir' from
ADA_RTL_OBJ_DIR from gcc/ada/gcc-interface/Makefile.in.
* configure: Regenerate.

From-SVN: r276424

libada: Remove racy duplicate gnatlib installation

For some reason, presumably historical, the `install-gnatlib' target for
the default multilib is invoked twice, once via the `ada.install-common'
target in `gcc/ada/gcc-interface/Make-lang.in' invoked from gcc/ and
again via the `install-libada' target in libada/.

Apart from doing the same twice this is actually harmful in sufficiently
parallelized `make' invocation, as the removal of old files performed
within the `install-gnatlib' recipe in the former case actually races
with the installation of new files done in the latter case, causing the
recipe to fail and abort, however non-fatally, having not completed the
installation of all the built files needed for the newly-built compiler
to work correctly.

This can be observed with a native `x86_64-linux-gnu' bootstrap:

make[4]: Entering directory '.../gcc/ada'
rm -rf .../lib/gcc/x86_64-linux-gnu/10.0.0/adalib
rm: cannot remove '.../lib/gcc/x86_64-linux-gnu/10.0.0/adalib': Directory not empty
make[4]: *** [gcc-interface/Makefile:512: install-gnatlib] Error 1
make[4]: Leaving directory '.../gcc/ada'
make[3]: *** [.../gcc/ada/gcc-interface/Make-lang.in:853: install-gnatlib] Error 2
make[2]: [.../gcc/ada/gcc-interface/Make-lang.in:829: ada.install-common] Error 2 (ignored)

which then causes missing files to be reported when an attempt is made
to use the newly-installed non-functional compiler to build a
`riscv-linux-gnu' cross-compiler:

(cd ada/bldtools/sinfo; gnatmake -q xsinfo ; ./xsinfo sinfo.h )
error: "ada.ali" not found, "ada.ads" must be compiled
error: "s-memory.ali" not found, "s-memory.adb" must be compiled
gnatmake: *** bind failed.
/bin/sh: ./xsinfo: No such file or directory
make[2]: *** [.../gcc/ada/Make-generated.in:45: ada/sinfo.h] Error 127
make[2]: Leaving directory '.../gcc'
make[1]: *** [Makefile:4369: all-gcc] Error 2
make[1]: Leaving directory '...'
make: *** [Makefile:965: all] Error 2

Depending on timing `.../lib/gcc/x86_64-linux-gnu/10.0.0/adainclude' may
cause an installation failure instead and the resulting compiler may be
non-functional in a different way.

Only invoke `install-gnatlib' from within gcc/ then if a legacy build
process is being used with libada disabled and gnatlib built manually
with `make -C gcc gnatlib'.

gcc/
* Makefile.in (gnat_install_lib): New variable.
* configure.ac: Substitute it.
* configure: Regenerate.

gcc/ada/
* gcc-interface/Make-lang.in (ada.install-common): Split into...
(gnat-install-tools, gnat-install-lib): ... these.

From-SVN: r276422

re PR c++/91222 (507.cactuBSSN_r build fails in warn_types_mismatch at ipa-devirt.c:1006 since r273571)

PR lto/91222
* ipa-devirt.c (warn_types_mismatch): Do not ICE when anonymous type
is matched with non-C++ type
* g++.dg/lto/odr-6_0.C: New testcase.
* g++.dg/lto/odr-6_1.c: New testcase.

From-SVN: r276420

ssa-thread-12.c: Fix warning introduced by my previous change.

* gcc.dg/tree-ssa/ssa-thread-12.c: Fix warning introduced by my
previous change.

From-SVN: r276418

tree-if-conv.c (tree_if_conversion): Move call to ifcvt_local_dce after local CSE.

2019-10-01  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>

* tree-if-conv.c (tree_if_conversion): Move call to ifcvt_local_dce
after local CSE.

From-SVN: r276417

invoke.texi (early-inlining-insns-O2): Document.

* doc/invoke.texi (early-inlining-insns-O2): Document.
(early-inlining-insns): Update.
* params.def (early-inlining-insns-O2): New bound.
(early-inlining-insns): Update docs.
* ipa-inline.c (want_early_inline_function_p): Use new bound.

* g++.dg/tree-ssa/pr61034.C: Set early-inlining-insns-O2=14.
* g++.dg/tree-ssa/pr8781.C: Likewise.
* g++.dg/warn/Wstringop-truncation-1.C: Likewise.
* gcc.dg/ipa/pr63416.c: likewise.
* gcc.dg/vect/pr66142.c: Likewise.
* gcc.dg/tree-ssa/ssa-thread-12.c: Mark compure_idf inline.

From-SVN: r276416

re PR c++/91925 (-fpack-struct causes a decltype with template to ICE)

PR c++/91925
* c-warn.c (check_alignment_of_packed_member): Ignore FIELD_DECLs
with NULL DECL_FIELD_OFFSET.

* g++.dg/conversion/packed2.C: New test.

From-SVN: r276415

re PR c++/88562 (Incorrect pointer incrementing on SH4)

gcc/
2019-10-01  Oleg Endo  <olegendo@gcc.gnu.org>

PR target/88562
* config/sh/sh.c (sh_extending_set_of_reg::use_as_extended_reg): Use
sh_check_add_incdec_notes to preserve REG_INC notes when replacing
a memory access insn.

From-SVN: r276411

rs6000-p8swap.c (rtx_is_swappable_p): Don't swap vpmsumd.

[gcc]

2019-10-01  Bill Schmidt  <wschmidt@linux.ibm.com>

* config/rs6000/rs6000-p8swap.c (rtx_is_swappable_p): Don't swap
vpmsumd.

[gcc/testsuite]

2019-10-01  Bill Schmidt  <wschmdit@linux.ibm.com>

* gcc.target/powerpc/pr91275.c: New.

From-SVN: r276410

S/390: Remove code duplication in vec_* comparison expanders

s390.md uses a lot of near-identical expanders that perform dispatching
to other expanders based on operand types. Since the following patch
would require even more of these, avoid copy-pasting the code by
generating these expanders using an iterator.

gcc/ChangeLog:

2019-10-01  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/77918
* config/s390/s390.c (s390_expand_vec_compare): Use
gen_vec_cmpordered and gen_vec_cmpunordered.
* config/s390/vector.md (vec_cmpuneq, vec_cmpltgt, vec_ordered,
vec_unordered): Delete.
(vec_ordered<mode>): Rename to vec_cmpordered<mode>.
(vec_unordered<mode>): Rename to vec_cmpunordered<mode>.
(VEC_CMP_EXPAND): New iterator for the generic dispatcher.
(vec_cmp<code>): Generic dispatcher.

From-SVN: r276409

S/390: Implement vcond expander for V1TI,V1TF

Currently gcc does not emit wf{c,k}* instructions when comparing long
double values.  Middle-end actually adds them in the first place, but
then veclower pass replaces them with floating point register pair
operations, because the corresponding expander is missing.

gcc/ChangeLog:

2019-10-01  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/77918
* config/s390/vector.md (V_HW): Add V1TI in order to make
vcond$a$b generate vcondv1tiv1tf.

From-SVN: r276408

Fix reload after function-abi patches (PR91948)

The code was passing a pseudo rather than its allocated hard reg
to ira_need_caller_save_p.  Running under valgrind to reproduce
the failure also showed that ALLOCNO_CROSSED_CALLS_ABIS wasn't
being explicitly initialised.

2019-10-01  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR rtl-optimization/91948
* ira-build.c (ira_create_allocno): Initialize
ALLOCNO_CROSSED_CALLS_ABIS.
* ira-color.c (allocno_reload_assign): Pass hard_regno rather
than regno to ira_need_caller_save_p.

From-SVN: r276407

recompute opt flags after opt level change

flag_omit_frame_pointer is set in machine-independent code depending
on the optimization level.  It is then overridden in x86
target-specific code depending on a macro defined by
--enable-frame-pointer.

Uses of attribute optimize go through machine-independent overriding
of flag_omit_frame_pointer, but the x86-specific overriding code did
NOT cover this flag, so, even if the attribute does not change the
optimization level, flag_omit_frame_pointer may end up with a
different value, and prevent inlining because of incompatible flags,
as detected by the gcc.dg/ipa/iinline-attr.c test on an
--enable-frame-pointer x86 toolchain.

for  gcc/ChangeLog

* config/i386/i386-options.c
(ix86_recompute_optlev_based_flags): New, moved out of...
(ix86_option_override_internal): ... this.  Call it.
(ix86_override_options_after_change): Call it here too.

From-SVN: r276405

Store float for pow result test

Optimizing gcc.dg/torture/pr41094.c, the compiler computes the
constant value and short-circuits the whole thing.  At -O0, however,
on 32-bit x86, the call to pow() remains, and the program compares the
returned value in a stack register, with excess precision, with the
exact return value expected from pow().  If libm's pow() returns a
slightly off result, the compare fails.  If the value in the register
is stored in a separate variable, so it gets rounded to double
precision, and then compared, the compare passes.

It's not clear that the test was meant to detect libm's reliance on
rounding off the excess precision, but I guess it wasn't, so I propose
this slight change that enables it to pass regardless of the slight
inaccuracy of the C library in use.

for  gcc/testsuite/ChangeLog

* gcc.dg/torture/pr41094.c: Introduce intermediate variable.

From-SVN: r276404

DWARF array bounds missing from C++ array definitions

A variable redeclaration or definition that provides additional type
information for it, e.g. outermost array bounds, is not reflected in
the debug information for the variable.  With this patch, the debug
info of the variable specialization gets a type attribute with the
adjusted type.

This patch affects mostly only array bounds.  However, when the
symbolic type used in a declaration and in a definition are different,
although they refer to the same type, debug information will end up
(correctly?) naming different symbolic types in the specification and
the definition.  Also, when a readonly declaration of an array loses
the readonly flag at the definition because of the initializer, the
definition may end up referencing a type while the specification
refers to a const-qualified version of that type.  If the type of the
variable is already const-qualified, e.g. an array of a const type,
the difference is meaningless.

for  gcc/ChangeLog

PR debug/91507
* dwarf2out.c (override_type_for_decl_p): New.
(gen_variable_die): Use it.

for  gcc/testsuite/ChangeLog

PR debug/91507
* gcc.dg/debug/dwarf2/array-0.c: New.
* gcc.dg/debug/dwarf2/array-1.c: New.
* gcc.dg/debug/dwarf2/array-2.c: New.
* gcc.dg/debug/dwarf2/array-3.c: New.
* g++.dg/debug/dwarf2/array-0.C: New.
* g++.dg/debug/dwarf2/array-1.C: New.
* g++.dg/debug/dwarf2/array-2.C: New.  Based on libstdc++-v3's
src/c++98/pool_allocator.cc:__pool_alloc_base::_S_heap_size.
* g++.dg/debug/dwarf2/array-3.C: New.  Based on
gcc's config/i386/i386-features.c:xlogue_layout::s_instances.
* g++.dg/debug/dwarf2/array-4.C: New.

From-SVN: r276403

tree-vect-loop.c (vectorizable_reduction): Move variables to where they are used.

2019-10-01  Richard Biener  <rguenther@suse.de>

* tree-vect-loop.c (vectorizable_reduction): Move variables
to where they are used.

From-SVN: r276402

regrename: Use PC instead of CC0 to hide operands

The regrename pass temporarily changes some operand RTL to CC0 so that
note_stores and scan_rtx don't see those operands.  CC0 is deprecated
and we want to remove it, so we need to use something else here.
PC fits the bill fine.

* regrename.c (hide_operands): Use pc_rtx instead of cc0_rtx.
(build_def_use): Use PC instead of CC0 in a comment.

From-SVN: r276401

Add myself to MAINTAINERS file

2019-10-01  Frederik Harwath <frederik@codesourcery.com>

 * MAINTAINERS: Add myself to Write After Approval

From-SVN: r276396

[C] Avoid aka types that just add tags

diag-aka-1.c tests that:

  struct T { int i; } T;
  void *a;
  T *t = a;

produces:

  request for implicit conversion from 'void *' to 'T *' {aka 'struct T *'} ...

But printing an aka for the tag seems a bit redundant when the tag name
is the same as the typedef name.  It's probably not going to be telling
the user anything they don't already know, and can be distracting if "T"
rather than "struct T" is the preferred choice for an exported interface.
This is even more true if the tag is anonymous; e.g.:

  struct { int i; } T;
  void *a;
  T *t = a;

gives:

  request for implicit conversion from 'void *' to 'T *' {aka 'struct <anonymous> *'}

Rather than just drop the test above, the patch instead tests for:

  struct T { int i; } *T;

where seeing the tag definitely helps.

2019-10-01  Richard Sandiford  <richard.sandiford@arm.com>

gcc/c/
* c-objc-common.c (useful_aka_type_p): New function.
(print_type): Use it to decide whether an aka type is worth printing.

gcc/testsuite/
* gcc.dg/diag-aka-1.c (T): Turn into a pointer typedef.
(foo): Update accordingly.
* gcc.dg/diag-aka-4.c: New test.

From-SVN: r276395

[C] Improve diagnostics for vector types

Given the following invalid arm_neon.h-based code:

  float x;
  int8x8_t y = x;

the error message we emit is pretty good:

  incompatible types when initializing type 'int8x8_t' using type 'float'

But convert the types to pointers:

  int8x8_t *ptr = &x;

and the message becomes:

  initialization of '__vector(8) signed char *' from incompatible pointer type 'float *'

Although it's reasonably obvious what '__vector(8) signed char *' means,
it isn't valid C or C++ syntax and is quite far from what the user wrote,
so using 'int8x8_t *' would be better.

This patch therefore prints the type name of vectors that have one.
It's still OK to print the __vector syntax as an "aka", although I have
a follow-on patch to tweak this slightly for types defined in system
header files.  The follow-on patch also addresses the ??? in
gcc.target/aarch64/diag_aka_1.c.

The C++ test already passed, but it seemed worth including for
consistency.

2019-10-01  Richard Sandiford  <richard.sandiford@arm.com>

gcc/c-family/
* c-pretty-print.c (pp_c_specifier_qualifier_list): If a vector type
has a type name, use it in preference to the __vector syntax.

gcc/testsuite/
* gcc.dg/diag-aka-3.c: New test.
* gcc.target/aarch64/diag_aka_1.c: New test.
* g++.dg/diagnostic/aka4.C: New test.

From-SVN: r276394

Remove clobber_high

The AArch64 SVE tlsdesc patterns were the main motivating reason
for clobber_high.  It's no longer needed now that the patterns use
calls instead.

At the time, one of the possible future uses for clobber_high was for
asm statements.  However, the current code wouldn't handle that case
without modification, so I think we might as well remove it for now.
We can always reapply it in future if it turns out to be useful again.

2019-10-01  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* rtl.def (CLOBBER_HIGH): Delete.
* doc/rtl.texi (clobber_high): Remove documentation.
* rtl.h (SET_DEST): Remove CLOBBER_HIGH from the list of codes.
(reg_is_clobbered_by_clobber_high): Delete.
(gen_hard_reg_clobber_high): Likewise.
* alias.c (record_set): Remove CLOBBER_HIGH handling.
* cfgexpand.c (expand_gimple_stmt): Likewise.
* combine-stack-adj.c (single_set_for_csa): Likewise.
* combine.c (find_single_use_1, set_nonzero_bits_and_sign_copies)
(can_combine_p, is_parallel_of_n_reg_sets, try_combine)
(record_dead_and_set_regs_1, reg_dead_at_p_1): Likewise.
* cse.c (invalidate_reg): Remove clobber_high parameter.
(invalidate): Update call accordingly.
(canonicalize_insn): Remove CLOBBER_HIGH handling.
(invalidate_from_clobbers, invalidate_from_sets_and_clobbers)
(count_reg_usage, insn_live_p): Likewise.
* cselib.h (cselib_invalidate_rtx): Remove sett argument.
* cselib.c (cselib_invalidate_regno, cselib_invalidate_rtx): Likewise.
(cselib_invalidate_rtx_note_stores): Update call accordingly.
(cselib_expand_value_rtx_1): Remove CLOBBER_HIGH handling.
(cselib_invalidate_regno, cselib_process_insn): Likewise.
* dce.c (deletable_insn_p, mark_nonreg_stores_1): Likewise.
(mark_nonreg_stores_2): Likewise.
* df-scan.c (df_find_hard_reg_defs, df_uses_record): Likewise.
(df_get_call_refs): Likewise.
* dwarf2out.c (mem_loc_descriptor): Likewise.
* emit-rtl.c (verify_rtx_sharing): Likewise.
(copy_insn_1, copy_rtx_if_shared_1): Likewise.
(hard_reg_clobbers_high, gen_hard_reg_clobber_high): Delete.
* genconfig.c (walk_insn_part): Remove CLOBBER_HIGH handling.
* genemit.c (gen_exp, gen_insn): Likewise.
* genrecog.c (validate_pattern, remove_clobbers): Likewise.
* haifa-sched.c (haifa_classify_rtx): Likewise.
* ira-build.c (create_insn_allocnos): Likewise.
* ira-costs.c (scan_one_insn): Likewise.
* ira.c (equiv_init_movable_p, memref_referenced_p): Likewise.
(rtx_moveable_p, interesting_dest_for_shprep): Likewise.
* jump.c (mark_jump_label_1): Likewise.
* lra-int.h (lra_insn_reg::clobber_high): Delete.
* lra-eliminations.c (lra_eliminate_regs_1): Remove CLOBBER_HIGH
handling.
(mark_not_eliminable): Likewise.
* lra-lives.c (process_bb_lives): Likewise.
* lra.c (new_insn_reg): Remove clobber_high parameter.
(collect_non_operand_hard_regs): Likewise.  Update call to new
insn_reg.  Remove CLOBBER_HIGH handling.
(lra_set_insn_recog_data): Remove CLOBBER_HIGH handling.  Update call
to collect_non_operand_hard_regs.
(add_regs_to_insn_regno_info): Remove CLOBBER_HIGH handling.
Update call to new_insn_reg.
(lra_update_insn_regno_info): Remove CLOBBER_HIGH handling.
* postreload.c (reload_cse_simplify, reload_combine_note_use)
(move2add_note_store): Likewise.
* print-rtl.c (print_pattern): Likewise.
* recog.c (store_data_bypass_p_1, store_data_bypass_p): Likewise.
(if_test_bypass_p): Likewise.
* regcprop.c (kill_clobbered_value, kill_set_value): Likewise.
* reginfo.c (reg_scan_mark_refs): Likewise.
* reload1.c (maybe_fix_stack_asms, eliminate_regs_1): Likewise.
(elimination_effects, mark_not_eliminable, scan_paradoxical_subregs)
(forget_old_reloads_1): Likewise.
* reorg.c (find_end_label, try_merge_delay_insns, redundant_insn)
(own_thread_p, fill_simple_delay_slots, fill_slots_from_thread)
(dbr_schedule): Likewise.
* resource.c (update_live_status, mark_referenced_resources)
(mark_set_resources): Likewise.
* rtl.c (copy_rtx): Likewise.
* rtlanal.c (reg_referenced_p, set_of_1, single_set_2, noop_move_p)
(note_pattern_stores): Likewise.
(reg_is_clobbered_by_clobber_high): Delete.
* sched-deps.c (sched_analyze_reg, sched_analyze_insn): Remove
CLOBBER_HIGH handling.

From-SVN: r276393

[AArch64] Use calls for SVE TLSDESC

One (unintended) side effect of the patches to support multiple
ABIs is that we can now represent tlsdesc calls as normal calls
on SVE targets.  This is likely to be handled more efficiently than
clobber_high, and for example fixes the long-standing failure in
gcc.target/aarch64/sve/tls_preserve_1.c.

2019-10-01  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR target/91452
* config/aarch64/aarch64.h (ARM_PCS_TLSDESC): New arm_pcs.
* config/aarch64/aarch64-protos.h (aarch64_tlsdesc_abi_id): Declare.
* config/aarch64/aarch64.c (aarch64_hard_regno_call_part_clobbered):
Handle ARM_PCS_TLSDESC.
(aarch64_tlsdesc_abi_id): New function.
* config/aarch64/aarch64.md (tlsdesc_small_sve_<mode>): Use a call
rtx instead of a list of clobbers and clobber_highs.
(tlsdesc_small_<mode>): Update accordingly.

From-SVN: r276392

[AArch64] Make call insns record the callee's arm_pcs

At the moment we rely on SYMBOL_REF_DECL to get the ABI of the callee
of a call insn, falling back to the default ABI if the decl isn't
available.  I think it'd be cleaner to attach the ABI directly to the
call instruction instead, which would also have the very minor benefit
of handling indirect calls more efficiently.

2019-10-01  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-protos.h (aarch64_expand_call): Take an
extra callee_abi argument.
* config/aarch64/aarch64.c (aarch64_expand_call): Likewise.
Insert a CALLEE_ABI unspec into the call pattern as the second
element in the PARALLEL.
(aarch64_simd_call_p): Delete.
(aarch64_insn_callee_abi): Get the arm_pcs of the callee from
the new CALLEE_ABI element of the PARALLEL.
(aarch64_init_cumulative_args): Get the arm_pcs of the callee
from the function type, if given.
(aarch64_function_arg_advance): Handle ARM_PCS_SIMD.
(aarch64_function_arg): Likewise.  Return the arm_pcs of the callee
when passed the function_arg_info end marker.
(aarch64_output_mi_thunk): Pass the arm_pcs of the callee as the
final argument of gen_sibcall.
* config/aarch64/aarch64.md (UNSPEC_CALLEE_ABI): New unspec.
(call): Make operand 2 a const_int_operand and pass it to expand_call.
Wrap it in an UNSPEC_CALLEE_ABI unspec for the dummy define_expand
pattern.
(call_value): Likewise operand 3.
(sibcall): Likewise operand 2.  Place the unspec before rather than
after the return.
(sibcall_value): Likewise operand 3.
(*call_insn, *call_value_insn): Include an UNSPEC_CALLEE_ABI.
(tlsgd_small_<mode>, *tlsgd_small_<mode>): Likewise.
(*sibcall_insn, *sibcall_value_insn): Likewise.  Remove empty
constraint strings.
(untyped_call): Pass const0_rtx as the callee ABI to gen_call.

gcc/testsuite/
* gcc.target/aarch64/torture/simd-abi-10.c: New test.
* gcc.target/aarch64/torture/simd-abi-11.c: Likewise.

From-SVN: r276391

configure.ac: Remove GCC_HEADER_STDINT(gstdint.h).

* configure.ac: Remove GCC_HEADER_STDINT(gstdint.h).
* libgomp.h: Include <stdint.h> instead of "gstdint.h".
* oacc-parallel.c: Don't include "libgomp_g.h".
* plugin/plugin-hsa.c: Include <stdint.h> instead of "gstdint.h".
* plugin/plugin-nvptx.c: Don't include "gstdint.h".
* aclocal.m4: Regenerated.
* config.h.in: Regenerated.
* configure: Regenerated.
* Makefile.in: Regenerated.

From-SVN: r276389

Update remaining calls to choose_hard_reg_mode

2019-10-01  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* regs.h (HARD_REGNO_CALLER_SAVE_MODE): Update call to
choose_hard_reg_mode.
* config/sparc/sparc.h (HARD_REGNO_CALLER_SAVE_MODE): Likewise.

From-SVN: r276388

doc/md.texi: Fix some typos

It says "size N/2" in a few places where "size S/2" is meant.

* doc/md.texi (vec_pack_trunc_@var{m}): Fix typo.
(vec_pack_sfix_trunc_@var{m}, vec_pack_ufix_trunc_@var{m}): Ditto.
(vec_packs_float_@var{m}, vec_packu_float_@var{m}): Ditto.

From-SVN: r276387

Daily bump.

From-SVN: r276386

compiler: change escape maps to hash tables

    Also use just one table lookup, not two.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/197759

From-SVN: r276382

* es.po: Update.

From-SVN: r276380

Implement C++20 constexpr comparison operators for __debug::array (P1023).

Add missing ChangeLog entry.

From-SVN: r276376

Implement C++20 constexpr comparison operators for __debug::array (P1023).

* include/debug/array: Add C++20 constexpr to comparison operators.
* testsuite/23_containers/array/tuple_interface/get_debug_neg.cc: Adapt
dg-error line numbers.
* testsuite/23_containers/array/tuple_interface/
tuple_element_debug_neg.cc: Likewise.

From-SVN: r276375

Include netinet/in.h in include/experimental/internet

    2019-09-30  Andreas Tobler  <andreast@gcc.gnu.org>

* include/experimental/internet: Include netinet/in.h if we have
_GLIBCXX_HAVE_NETINET_IN_H defined.

From-SVN: r276374

libstdc++ testsuite, silence a FreeBSD libm warning

    Add missing ChangeLog entry for the previous commit.

From-SVN: r276373

libstdc++ testsuite, silence a FreeBSD libm warning

    2019-09-30  Andreas Tobler  <andreast@gcc.gnu.org>

* testsuite/ext/special_functions/airy_ai/check_nan.cc: Ignore the
FreeBSD warning about lower advertised precision of tgammal.
* testsuite/ext/special_functions/airy_bi/check_nan.cc: Likewise.
* testsuite/special_functions/07_cyl_bessel_i/check_nan.cc: Likewise.
* testsuite/special_functions/08_cyl_bessel_j/check_nan.cc: Likewise.
* testsuite/special_functions/09_cyl_bessel_k/check_nan.cc: Likewise.
* testsuite/special_functions/10_cyl_neumann/check_nan.cc: Likewise.
* testsuite/special_functions/19_sph_bessel/check_nan.cc: Likewise.
* testsuite/special_functions/21_sph_neumann/check_nan.cc: Likewise.
* testsuite/tr1/5_numerical_facilities/special_functions/
08_cyl_bessel_i/check_nan.cc: Likewise.
* testsuite/tr1/5_numerical_facilities/special_functions/
09_cyl_bessel_j/check_nan.cc: Likewise.
* testuite/tr1/5_numerical_facilities/special_functions/
10_cyl_bessel_k/check_nan.cc: Likewise.
* testsuite/tr1/5_numerical_facilities/special_functions/
11_cyl_neumann/check_nan.cc: Likewise.
* testsuite/tr1/5_numerical_facilities/special_functions/
21_sph_bessel/check_nan.cc: Likewise.
* testsuite/tr1/5_numerical_facilities/special_functions/
23_sph_neumann/check_nan.cc: Likewise.

From-SVN: r276372

diagnostic-show-locus.c: rework handling of multiple labels

This patch improves the handling of large numbers of labels within a
rich_location: previously, overlapping labels could lead to an assertion
failure within layout::print_any_labels.  Also, the labels were printed
in reverse order of insertion into the rich_location.

This patch moves the determination of whether a vertical bar should
be printed for a line_label into the
  'Figure out how many "label lines" we need, and which
   one each label is printed in.'
step of layout::print_any_labels, rather than doing it as the lines
are printed.  It also flips the sort order, so that labels at the
same line/column are printed in order of insertion into the
rich_location.

I haven't run into these issues with our existing diagnostics, but it
affects a patch kit I'm working on that makes more extensive use of
labels.

gcc/ChangeLog:
* diagnostic-show-locus.c (line_label::line_label): Initialize
m_has_vbar.
(line_label::comparator): Reverse the sort order by m_state_idx,
so that when the list is walked backwards the labels appear in
order of insertion into the rich_location.
(line_label::m_has_vbar): New field.
(layout::print_any_labels): When dealing with multiple labels at
the same line and column, only print vertical bars for the one
with the highest label_line.
(selftest::test_one_liner_labels): Update test for multiple labels
to expect the labels to be in the order of insertion into the
rich_location.  Add a test for many such labels, where the column
numbers are out-of-order relative to the insertion order.

From-SVN: r276371

[x86] Cache result of expensive_function_p between frame layouts

ix86_compute_frame_layout sets use_fast_prologue_epilogue if
the function isn't more expensive than a certain threshold,
where the threshold depends on the number of saved registers.
However, the RA is allowed to insert and delete instructions
as it goes along, which can change whether this threshold is
crossed or not.

I hit this with an RA change I'm working on.  Rematerialisation
was able to remove an instruction and avoid a spill, which happened
to bring the size of the function below the threshold.  But since
nothing legitimately frame-related had changed, there was no need for
the RA to lay out the frame again.  We then failed the final sanity
check in lra_eliminate.

2019-09-30  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/i386/i386.h (ix86_frame::expensive_p): New field.
(ix86_frame::expensive_count): Likewise.
* config/i386/i386.c (ix86_compute_frame_layout): Make the choice
of use_fast_prologue_epilogue robust against incidental changes
in function size.

From-SVN: r276361

S/390: Remove code duplication in vec_unordered<mode>

vec_unordered<mode> is vec_ordered<mode> plus a negation at the end.
Reuse vec_unordered<mode> logic.

gcc/ChangeLog:

2019-09-30  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/77918
* config/s390/vector.md (vec_unordered<mode>): Call
gen_vec_ordered<mode>.

From-SVN: r276360

Fix typo in my last commit's ChangeLog entry

From-SVN: r276359

[AArch64][SVE] Utilize ASRD instruction for division and remainder

2019-09-30  Yuliang Wang  <yuliang.wang@arm.com>

gcc/
* config/aarch64/aarch64-sve.md (sdiv_pow2<mode>3):
New pattern for ASRD.
* config/aarch64/iterators.md (UNSPEC_ASRD): New unspec.
* internal-fn.def (IFN_DIV_POW2): New internal function.
* optabs.def (sdiv_pow2_optab): New optab.
* tree-vect-patterns.c (vect_recog_divmod_pattern):
Modify pattern to support new operation.
* doc/md.texi (sdiv_pow2$var{m3}): Documentation for the above.
* doc/sourcebuild.texi (vect_sdiv_pow2_si):
Document new target selector.

gcc/testsuite/
* gcc.dg/vect/vect-sdiv-pow2-1.c: New test.
* gcc.target/aarch64/sve/asrdiv_1.c: As above.
* lib/target-supports.exp (check_effective_target_vect_sdiv_pow2_si):
Return true for AArch64 with SVE.

From-SVN: r276343

[AArch64] Make more use of function_abi

This patch makes more use of the function_abi infrastructure.
We can then avoid checking specifically for the vector PCS in
a few places, and can test it more directly otherwise.

Specifically: we no longer need to call df_set_regs_ever_live
for the extra call-saved registers, since IRA now does that for us.
We also don't need to handle the vector PCS specially in
aarch64_epilogue_uses, because DF now marks the registers
as live on exit.

2019-09-30  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.c (aarch64_layout_frame): Use crtl->abi
to test whether we're compiling a vector PCS function and to test
whether the function needs to save a particular register.
Remove the vector PCS handling of df_set_regs_ever_live.
(aarch64_components_for_bb): Use crtl->abi to test whether
the function needs to save a particular register.
(aarch64_process_components): Use crtl->abi to test whether
we're compiling a vector PCS function.
(aarch64_expand_prologue, aarch64_expand_epilogue): Likewise.
(aarch64_epilogue_uses): Remove handling of vector PCS functions.

From-SVN: r276341

[AArch64] Allow shrink-wrapping of non-leaf vector PCS functions

With the function ABI stuff, we can now support shrink-wrapping of
non-leaf vector PCS functions.  This is particularly useful if the
vector PCS function calls an ordinary function on an error path,
since we can then keep the extra saves and restores specific to
that path too.

2019-09-30  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-protos.h (aarch64_use_simple_return_insn_p):
Delete.
* config/aarch64/aarch64.c (aarch64_components_for_bb): Check
whether the block calls a function that clobbers more registers
than the current function is allowed to.
(aarch64_use_simple_return_insn_p): Delete.
* config/aarch64/aarch64.md (simple_return): Remove condition.

gcc/testsuite/
* gcc.target/aarch64/torture/simd-abi-9.c: New test.

From-SVN: r276340

Make ira call df_set_regs_ever_live for extra call-clobbered regs

If we support multiple ABIs in the same translation unit, it can
sometimes be the case that a callee clobbers more registers than
its caller is allowed to.  We need to call df_set_regs_ever_live
on these extra registers so that the prologue and epilogue code
can handle them appropriately.

This patch does that in IRA.  I wanted to avoid another full
instruction walk just for this, so I combined it with the existing
set_paradoxical_subreg walk.  This happens before the first
calculation of elimination offsets.

2019-09-30  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* function-abi.h (function_abi_aggregator): New class.
* function-abi.cc (function_abi_aggregator::caller_save_regs): New
function.
* ira.c (update_equiv_regs_prescan): New function.  Call
set_paradoxical_subreg here rather than...
(update_equiv_regs): ...here.
(ira): Call update_equiv_regs_prescan.

From-SVN: r276339