compiler, runtime: provide index information on bounds check failure

    This implements https://golang.org/cl/161477 in the gofrontend.

    Updates golang/go#30116

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/191881

From-SVN: r274998

PR tree-optimization/91457 - inconsistent warning for writing past the end of an array member

gcc/ChangeLog:

PR tree-optimization/91457
* builtins.c (component_size): New function.
(compute_objsize): Add argument. Handle ARRAY_REF and COMPONENT_REF.
* builtins.h (compute_objsize): Add argument.
* tree-ssa-strlen.c (handle_store): Handle no-warning bit.
* tree-vrp.c (vrp_prop::check_array_ref): Return warning result.
(vrp_prop::check_mem_ref): Same.
(vrp_prop::search_for_addr_array): Set no-warning bit.
(check_array_bounds): Same.

gcc/testsuite/ChangeLog:

PR tree-optimization/91457
* c-c++-common/Wstringop-overflow-2.c: New test.
* g++.dg/warn/Warray-bounds-8.C: New test.
* g++.dg/warn/Wstringop-overflow-3.C: New test.
* gcc.dg/Wstringop-overflow-15.c: New test.

From-SVN: r274997

PR driver/80545 - option -Wstringop-overflow not recognized by Fortran

gcc/ChangeLog:

PR driver/80545
* opts-common.c (option_enabled): Correct checking for language
options.

From-SVN: r274996

* config/i386/i386-features.c
(general_scalar_chain::compute_convert_gain):
Correct cost for double-word shifts.
(general_scalar_to_vector_candidate_p): Reject count operands
greater or equal to mode bitsize.

From-SVN: r274994

[PR 91468] Small fixes in ipa-cp.c and ipa-prop.c

2019-08-28  Martin Jambor  <mjambor@suse.cz>

PR ipa/91468
* ipa-cp.c (merge_agg_lats_step): Removed redundant test, made a
checking assert a normal assert to test it really is redundant.
* ipa-prop.c (compute_complex_assign_jump_func): Removed
redundant test.
(update_jump_functions_after_inlining): Removed combining unary
arithmetic operations with an ancestor jump function.
(ipcp_modif_dom_walker::before_dom_children): Fix wrong use of rhs
instead of t.

From-SVN: r274992

[PR c++/90613] Fix using-decl debug bloat

https://gcc.gnu.org/ml/gcc-patches/2019-08/msg01888.html
cp/
PR c++/90613
* name-lookup.c (cp_emit_debug_info): Check for builtins during
overload iteration.

testsuite/
PR c++/90613
* g++.dg/lookup/using61.C: New.

From-SVN: r274991

i386-features.c (convert_scalars_to_vector): Do not add the MD problem.

2019-08-28  Richard Biener  <rguenther@suse.de>

* config/i386/i386-features.c (convert_scalars_to_vector): Do not
add the MD problem.

From-SVN: r274990

Stated standards in documentation

Correct the stated standards in documentation and for intrinsics and
specific intrinsics.  In C_SIZEOF the printed value is T not .TRUE..
In IPARITY example wrap BOZ constants in calls to INT.

From-SVN: r274988

re PR middle-end/89544 (Argument marshalling incorrectly assumes stack slots are naturally aligned.)

2019-08-28  Bernd Edlinger  <bernd.edlinger@hotmail.de>

        PR middle-end/89544
        * gcc.target/arm/unaligned-argument-3.c: New test.

From-SVN: r274987

expr.c (expand_assignment): Handle misaligned DECLs.

2019-09-28  Bernd Edlinger  <bernd.edlinger@hotmail.de>
    Richard Biener  <rguenther@suse.de>

* expr.c (expand_assignment): Handle misaligned DECLs.
(expand_expr_real_1): Handle FUNCTION_DECL as unaligned.
* function.c (assign_parm_adjust_stack_rtl): Check movmisalign optab
too.
(assign_parm_setup_stack): Allocate properly aligned stack slots.
* varasm.c (build_constant_desc): Align constants of misaligned types.
* config/arm/predicates.md (aligned_operand): New predicate.
* config/arm/arm.md (movdi, movsi, movhi, movhf, movsf, movdf): Use
aligned_operand to check restrictions on memory addresses.
* config/arm/neon.md (movti, mov<VSTRUCT>, mov<VH>): Likewise.
* config/arm/vec-common.md (mov<VALL>): Likewise.

Co-Authored-By: Richard Biener <rguenther@suse.de>
From-SVN: r274986

re PR libgomp/91530 (Several libgomp.*/scan-* tests FAIL without avx_runtime)

PR libgomp/91530
* config/i386/sse.md (vec_shl_<mode>, vec_shr_<mode>): Use
V_128 iterator instead of VI_128.

* testsuite/libgomp.c/scan-21.c: New test.
* testsuite/libgomp.c/scan-22.c: New test.

From-SVN: r274985

re PR libgomp/91530 (Several libgomp.*/scan-* tests FAIL without avx_runtime)

PR libgomp/91530
* config/i386/sse.md (vec_shl_<mode>, vec_shr_<mode>): Use
V_128 iterator instead of VI_128.

* testsuite/libgomp.c/scan-21.c: New test.
* testsuite/libgomp.c/scan-22.c: New test.

From-SVN: r274984

Remove code leftover that has never been used.

2019-08-28  Martin Liska  <mliska@suse.cz>

PR tree-optimization/90970
* builtins.c (check_access): Remove assignment to maxread
as it hasn't been used since when it was introduced in r255755.

From-SVN: r274983

PR c++/81676 - bogus -Wunused warnings in constexpr if.

* semantics.c (maybe_mark_exp_read_r): New function.
(finish_if_stmt): Call it on THEN_CLAUSE and ELSE_CLAUSE.

* g++.dg/cpp1z/constexpr-if31.C: New test.
* g++.dg/cpp1z/constexpr-if32.C: New test.

From-SVN: r274982

PR c++/91428 - warn about std::is_constant_evaluated in if constexpr.

* cp-tree.h (decl_in_std_namespace_p): Declare.
* semantics.c (is_std_constant_evaluated_p): New.
(finish_if_stmt_cond): Warn about "std::is_constant_evaluated ()" in
an if-constexpr.
* typeck.c (decl_in_std_namespace_p): No longer static.

* g++.dg/cpp2a/is-constant-evaluated9.C: New test.

From-SVN: r274981

Daily bump.

From-SVN: r274980

PR tree-optimization/91567 - Spurious -Wformat-overflow warnings building glibc (32-bit only)

gcc/ChangeLog:

PR tree-optimization/91567
* gimple-ssa-sprintf.c (get_string_length): Handle more forms of lengths
of unknown strings.
* vr-values.c (vr_values::extract_range_basic): Set strlen upper bound
to PTRDIFF_MAX - 2.

gcc/testsuite/ChangeLog:

PR tree-optimization/91567
* gcc.dg/tree-ssa/builtin-snprintf-6.c: Xfail a subset of assertions
on targets other than x86_64 to work around PR 83543.
* gcc.dg/tree-ssa/builtin-sprintf-warn-22.c: New test.

From-SVN: r274976

tree-ssa-strlen.c (printf_strlen_execute): Initialize the loop optimizer and SCEV before sizing ssa_ver_to_stridx.

* tree-ssa-strlen.c (printf_strlen_execute): Initialize
the loop optimizer and SCEV before sizing ssa_ver_to_stridx.

* gcc.c-torture/compile/20190827-1.c: New test.

From-SVN: r274975

re PR fortran/91496 (!GCC$ directives error if mistyped or unknown)

2019-08-27  Harald Anlauf  <anlauf@gmx.de>

PR fortran/91496
* gfortran.h: Extend struct gfc_iterator for loop annotations.
* array.c (gfc_copy_iterator): Copy loop annotations by IVDEP,
VECTOR, and NOVECTOR pragmas.
* decl.c (gfc_match_gcc_ivdep, gfc_match_gcc_vector)
(gfc_match_gcc_novector): New matcher functions handling IVDEP,
VECTOR, and NOVECTOR pragmas.
* match.h: Declare prototypes of matcher functions handling IVDEP,
VECTOR, and NOVECTOR pragmas.
* parse.c (decode_gcc_attribute, parse_do_block)
(parse_executable): Decode IVDEP, VECTOR, and NOVECTOR pragmas;
emit warning for unrecognized pragmas instead of error.
* trans-stmt.c (gfc_trans_simple_do, gfc_trans_do): Add code to
emit annotations for IVDEP, VECTOR, and NOVECTOR pragmas.
* gfortran.texi: Document IVDEP, VECTOR, and NOVECTOR pragmas.

PR fortran/91496
* gfortran.dg/pr91496.f90: New testcase.

From-SVN: r274966

sse4_1-round-roundeven-1.c (dg-options): Add -mfpmath=sse.

* gcc.target/i386/sse4_1-round-roundeven-1.c (dg-options):
Add -mfpmath=sse.
* gcc.target/i386/sse4_1-round-roundeven-2.c (dg-options): Ditto.

From-SVN: r274964

re PR target/91528 (ICE in ix86_expand_prologue at i386.c:7844 since r274481)

PR target/91528
* config/i386/i386-features.c (convert_scalars_to_vector):
Update crtl->stack_realign_needed, crtl->stack_realign_tried and
crtl->stack_realign_processed.  Update crtl->drap_reg by calling
targetm.calls.get_drap_rtx.  If drap_rtx is non-null then
Update crtl->args.internal_arg_pointer and call fixup_tail_calls.

testsuite/ChangeLog:

PR target/91528
* gcc.target/i386/pr91528.c: New test.

From-SVN: r274962

re PR c++/83431 (-Wformat-truncation may incorrectly report truncation)

gcc/testsuite/ChangeLog:

PR c++/83431
PR testsuite/91562
* gcc.dg/strlenopt-8.c: Adjust pass/dump name.

From-SVN: r274961

Options documentation fixes.

Ensure that options lists fit cleanly inside the margins of a PDF page.
Reword description of option -ffrontend-loop-interchange so that it fits
cleanly inside the margins of a PDF page. Add options to those enabled
by -fdec.

From-SVN: r274958

libgo: rebuild runtime.inc if mkruntimeinc.sh changes

    The Makefile was missing a dependency.

    Also remove runtime.inc.raw in mostlyclean.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/191958

From-SVN: r274956

Share a prevailing name for remove debug info symbols w/ LTO.

2019-08-27  Martin Liska  <mliska@suse.cz>

PR lto/91478
* simple-object-elf.c (simple_object_elf_copy_lto_debug_sections):
First find a WEAK HIDDEN symbol in symbol table that will be
preserved.  Later, use the symbol name for all removed symbols.

From-SVN: r274955

2019-08-27  Richard Biener  <rguenther@suse.de>

* config/i386/i386-features.h
(general_scalar_chain::~general_scalar_chain): Add.
(general_scalar_chain::insns_conv): New bitmap.
(general_scalar_chain::n_sse_to_integer): New.
(general_scalar_chain::n_integer_to_sse): Likewise.
(general_scalar_chain::make_vector_copies): Adjust signature.
* config/i386/i386-features.c
(general_scalar_chain::general_scalar_chain): Outline,
initialize new members.
(general_scalar_chain::~general_scalar_chain): New.
(general_scalar_chain::mark_dual_mode_def): Record insns
we need to insert conversions at and count them.
(general_scalar_chain::compute_convert_gain): Account
for conversion instructions at chain boundary.
(general_scalar_chain::make_vector_copies): Generate a single
copy for a def by a specific insn.
(general_scalar_chain::convert_registers): First populate
defs_map, then make copies at out-of chain insns.

From-SVN: r274953

re PR c++/91415 (Invalid warning for C++17 sequencing of shift operator E1<<E2.)

PR c++/91415
* c-common.c (verify_tree): For LSHIFT_EXPR, RSHIFT_EXPR,
COMPONENT_REF and ARRAY_REF in cxx_dialect >= cxx17 mode handle it
like COMPOUND_EXPR rather than normal expression.

* g++.dg/warn/sequence-pt-4.C: New test.

From-SVN: r274952

re PR testsuite/91549 (gcc.dg/wrapped-binop-simplify.c fails starting with r274925)

PR testsuite/91549

gcc/testsuite/ChangeLog:
* gcc.dg/wrapped-binop-simplify.c: Test only on x86, s390 with lp64.

From-SVN: r274951

re PR libgomp/91530 (Several libgomp.*/scan-* tests FAIL without avx_runtime)

PR libgomp/91530
* testsuite/libgomp.c/scan-11.c: Add -msse2 option for sse2_runtime
targets.
* testsuite/libgomp.c/scan-12.c: Likewise.
* testsuite/libgomp.c/scan-13.c: Likewise.
* testsuite/libgomp.c/scan-14.c: Likewise.
* testsuite/libgomp.c/scan-15.c: Likewise.
* testsuite/libgomp.c/scan-16.c: Likewise.
* testsuite/libgomp.c/scan-17.c: Likewise.
* testsuite/libgomp.c/scan-18.c: Likewise.
* testsuite/libgomp.c/scan-19.c: Likewise.
* testsuite/libgomp.c/scan-20.c: Likewise.
* testsuite/libgomp.c++/scan-9.C: Likewise.
* testsuite/libgomp.c++/scan-10.C: Likewise.
* testsuite/libgomp.c++/scan-11.C: Likewise.
* testsuite/libgomp.c++/scan-12.C: Likewise.
* testsuite/libgomp.c++/scan-14.C: Likewise.
* testsuite/libgomp.c++/scan-15.C: Likewise.
* testsuite/libgomp.c++/scan-13.C: Likewise.  Use sse2_runtime
instead of i?86-*-* x86_64-*-* as target for scan-tree-dump-times.
* testsuite/libgomp.c++/scan-16.C: Likewise.

From-SVN: r274947

[arm/aarch64] Add comments warning that stack-protector initializer insns shouldn't be split

Following the publication of https://kb.cert.org/vuls/id/129209/ I've
been having a look at GCC's implementation for Arm and AArch64.  I
haven't identified any issues yet, but it's a bit early to be
completely sure.

One observation, however, is that the instruction sequence that
initializes the stack canary might be vulnerable to producing a
reusable value if it were ever split early.  I don't think we ever
would, because the memory locations involved with the stack protector
are all marked volatile to ensure that the values are only loaded at
the point in time when the test is intended to happen, and that also
has the effect of making it unlikely that the value would be reused
without reloading.  Nevertheless, defence in depth is probably
warranted here.

So this patch just adds some comments warning that the patterns should
not be split.

* config/arm/arm.md (stack_protect_set_insn): Add security-related
comment.
* config/aarch64/aarch64.md (stack_protect_set_<mode>): Likewise.

From-SVN: r274946

Fix new clang warnings.

2019-08-27  Martin Liska  <mliska@suse.cz>

* cgraph.c (cgraph_node::remove): Remove dead assignment before
loop.
* config/i386/i386-features.c (scalar_chain::emit_conversion_insns):
Enclose in anonymous namespace.
* config/i386/x86-tune-costs.h (struct processor_costs): Wrap
hard_register initialization in braces.
* tree-vrp.h (value_range_base::supports_type_p): Return false
for function with boolean return type.

From-SVN: r274945

Daily bump.

From-SVN: r274944

re PR fortran/91390 (treatment of extra parameter in a subroutine call)

2019-08-26  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/91390
PR fortran/91473
* frontend-passes.c (gfc_check_externals): Make
gfc_errors_to_warnings conditional on -fallow-argument-mismatch.
* invoke.texi: Document -fallow-argument-mismatch.
* lang.opt: Add -fallow-argument-mismatch.

2019-08-26  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/91390
PR fortran/91473
* gfortran.dg/used_before_typed_4.f90: Change warning to error.
* gfortran.dg/argument_checking_20.f90: New test.

From-SVN: r274937

compiler: generalize cleanup of unresolved placeholder pointer types

    This change extends the work in https://golang.org/cl/51131 to include
    placeholder pointer types created for Go function types, which can
    also be left dangling/unresolved in some instances. This fixes an
    assert in Llvm_backend::materializeComposite.

    Test case can be found in https://golang.org/cl/191743.

    Updates golang/go#33020.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/191744

From-SVN: r274935

i386.c (emit_i387_cw_initialization): Fix masking operand value.

* config/i386/i386.c (emit_i387_cw_initialization)
<case I387_CW_ROUNDEVEN>: Fix masking operand value.

From-SVN: r274934

PR tree-optimization/83431 - -Wformat-truncation may incorrectly report truncation

gcc/ChangeLog:

PR c++/83431
* gimple-ssa-sprintf.c (pass_data_sprintf_length): Remove object.
(sprintf_dom_walker): Remove class.
(get_int_range): Make argument const.
(directive::fmtfunc, directive::set_precision): Same.
(format_none): Same.
(build_intmax_type_nodes): Same.
(adjust_range_for_overflow): Same.
(format_floating): Same.
(format_character): Same.
(format_string): Same.
(format_plain): Same.
(get_int_range): Cast away constness.
(format_integer): Same.
(get_string_length): Call get_range_strlen_dynamic.  Handle
null lendata.maxbound.
(should_warn_p): Adjust argument scope qualifier.
(maybe_warn): Same.
(format_directive): Same.
(parse_directive): Same.
(is_call_safe): Same.
(try_substitute_return_value): Same.
(sprintf_dom_walker::handle_printf_call): Rename...
(handle_printf_call): ...to this.  Initialize target to host charmap
here instead of in pass_sprintf_length::execute.
(struct call_info): Make global.
(sprintf_dom_walker::compute_format_length): Make global.
(sprintf_dom_walker::handle_gimple_call): Same.
* passes.def (pass_sprintf_length): Replace with pass_strlen.
* print-rtl.c (print_pattern): Reduce the number of spaces to
avoid -Wformat-truncation.
* tree-pass.h (make_pass_warn_printf): New function.
* tree-ssa-strlen.c (strlen_optimize): New variable.
(get_string_length): Add comments.
(get_range_strlen_dynamic): New function.
(check_and_optimize_call): New function.
(handle_integral_assign): New function.
(strlen_check_and_optimize_stmt): Factor code out into
strlen_check_and_optimize_call and handle_integral_assign.
(strlen_dom_walker::evrp): New member.
(strlen_dom_walker::before_dom_children): Use evrp member.
(strlen_dom_walker::after_dom_children): Use evrp member.
(printf_strlen_execute): New function.
(pass_strlen::gate): Update to handle printf calls.
(dump_strlen_info): New function.
(pass_data_warn_printf): New variable.
(pass_warn_printf): New class.
* tree-ssa-strlen.h (get_range_strlen_dynamic): Declare.
(handle_printf_call): Same.
* tree-vrp.c (value_range_base::type): Adjust assertion.
* vr-values.c (vr_values::update_value_range): Use type of the first
argument rather than the second.

gcc/testsuite/ChangeLog:

PR c++/83431
* gcc.dg/strlenopt-63.c: New test.
* gcc.dg/pr79538.c: Adjust text of expected warning.
* gcc.dg/pr81292-1.c: Adjust pass name.
* gcc.dg/pr81292-2.c: Same.
* gcc.dg/pr81703.c: Same.
* gcc.dg/strcmpopt_2.c: Same.
* gcc.dg/strcmpopt_3.c: Same.
* gcc.dg/strcmpopt_4.c: Same.
* gcc.dg/strlenopt-1.c: Same.
* gcc.dg/strlenopt-10.c: Same.
* gcc.dg/strlenopt-11.c: Same.
* gcc.dg/strlenopt-13.c: Same.
* gcc.dg/strlenopt-14g.c: Same.
* gcc.dg/strlenopt-14gf.c: Same.
* gcc.dg/strlenopt-15.c: Same.
* gcc.dg/strlenopt-16g.c: Same.
* gcc.dg/strlenopt-17g.c: Same.
* gcc.dg/strlenopt-18g.c: Same.
* gcc.dg/strlenopt-19.c: Same.
* gcc.dg/strlenopt-1f.c: Same.
* gcc.dg/strlenopt-2.c: Same.
* gcc.dg/strlenopt-20.c: Same.
* gcc.dg/strlenopt-21.c: Same.
* gcc.dg/strlenopt-22.c: Same.
* gcc.dg/strlenopt-22g.c: Same.
* gcc.dg/strlenopt-24.c: Same.
* gcc.dg/strlenopt-25.c: Same.
* gcc.dg/strlenopt-26.c: Same.
* gcc.dg/strlenopt-27.c: Same.
* gcc.dg/strlenopt-28.c: Same.
* gcc.dg/strlenopt-29.c: Same.
* gcc.dg/strlenopt-2f.c: Same.
* gcc.dg/strlenopt-3.c: Same.
* gcc.dg/strlenopt-30.c: Same.
* gcc.dg/strlenopt-31g.c: Same.
* gcc.dg/strlenopt-32.c: Same.
* gcc.dg/strlenopt-33.c: Same.
* gcc.dg/strlenopt-33g.c: Same.
* gcc.dg/strlenopt-34.c: Same.
* gcc.dg/strlenopt-35.c: Same.
* gcc.dg/strlenopt-4.c: Same.
* gcc.dg/strlenopt-48.c: Same.
* gcc.dg/strlenopt-49.c: Same.
* gcc.dg/strlenopt-4g.c: Same.
* gcc.dg/strlenopt-4gf.c: Same.
* gcc.dg/strlenopt-5.c: Same.
* gcc.dg/strlenopt-50.c: Same.
* gcc.dg/strlenopt-51.c: Same.
* gcc.dg/strlenopt-52.c: Same.
* gcc.dg/strlenopt-53.c: Same.
* gcc.dg/strlenopt-54.c: Same.
* gcc.dg/strlenopt-55.c: Same.
* gcc.dg/strlenopt-56.c: Same.
* gcc.dg/strlenopt-6.c: Same.
* gcc.dg/strlenopt-61.c: Same.
* gcc.dg/strlenopt-7.c: Same.
* gcc.dg/strlenopt-8.c: Same.
* gcc.dg/strlenopt-9.c: Same.
* gcc.dg/strlenopt.h (snprintf, snprintf): Declare.
* gcc.dg/tree-ssa/builtin-snprintf-6.c: New test.
* gcc.dg/tree-ssa/builtin-snprintf-7.c: New test.
* gcc.dg/tree-ssa/builtin-snprintf-8.c: New test.
* gcc.dg/tree-ssa/builtin-snprintf-9.c: New test.
* gcc.dg/tree-ssa/builtin-sprintf-warn-21.c: New test.
* gcc.dg/tree-ssa/dump-4.c: New test.
* gcc.dg/tree-ssa/pr83501.c: Adjust pass name.

From-SVN: r274933

* decl.c (duplicate_decls): Always merge DECL_DECLARED_CONSTEXPR_P.

From-SVN: r274932

* ChangeLog: Fix roundeven entry.

From-SVN: r274931

PR c++/91545 - ICE in constexpr store evaluation.

* constexpr.c (cxx_eval_store_expression): Check FIELD_DECL instead
of DECL_P.

* g++.dg/cpp0x/pr91545.C: New test.

From-SVN: r274930

i386-features.c (general_remove_non_convertible_regs): Remove.

2019-08-26  Richard Biener  <rguenther@suse.de>

* config/i386/i386-features.c (general_remove_non_convertible_regs):
Remove.
(convert_scalars_to_vector): Do not call it.

From-SVN: r274929

i386: Roundeven expansion for SSE4.1+

gcc/ChangeLog:

2019-08-26  Tejas Joshi  <tejasjoshi9673@gmail.com>
            Uros Bizjak  <ubizjak@gmail.com>

* builtins.c (mathfn_built_in_2): Change CASE_MATHFN to
CASE_MATHFN_FLOATN for roundeven.
* config/i386/i386.c (ix86_i387_mode_needed): Add case
I387_ROUNDEVEN.
(ix86_mode_needed): Likewise.
(ix86_mode_after): Likewise.
(ix86_mode_entry): Likewise.
(ix86_mode_exit): Likewise.
(ix86_emit_mode_set): Likewise.
(emit_i387_cw_initialization): Add case I387_CW_ROUNDEVEN.
* config/i386/i386.h (ix86_stack_slot): Add SLOT_CW_ROUNDEVEN.
(ix86_entry): Add I387_ROUNDEVEN.
(avx_u128_state): Add I387_CW_ANY.
* config/i386/i386.md: Define UNSPEC_FRNDINT_ROUNDEVEN.
(define_int_iterator): Likewise.
(define_int_attr): Likewise for rounding_insn, rounding and ROUNDING.
(define_constant): Define ROUND_ROUNDEVEN mode.
(define_attr): Add roundeven mode for i387_cw.
(<rouding_insn><mode>2): Add condition for ROUND_ROUNDEVEN.
* internal-fn.def (ROUNDEVEN): New builtin function.
* optabs.def (roundeven_optab): New optab.

gcc/testsuite/ChangeLog:

2019-08-26  Tejas Joshi  <tejasjoshi9673@gmail.com>

* gcc.target/i386/sse4_1-round-roundeven-1.c: New test.
* gcc.target/i386/sse4_1-round-roundeven-2.c: New test.

Co-Authored-By: Uros Bizjak <ubizjak@gmail.com>
From-SVN: r274928

Builtin function roundeven folding implementation

2019-08-26  Tejas Joshi  <tejasjoshi9673@gmail.com>

        * builtins.c (mathfn_built_in_2): Added CASE_MATHFN_FLOATN
        for ROUNDEVEN.
        * builtins.def: Added function definitions for roundeven function
        variants.
        * fold-const-call.c (fold_const_call_ss): Added case for roundeven
        function call.  Adjust condition for floor, ceil, trunc and round.
        * fold-const.c (negate_mathfn_p): Added case for roundeven function.
        (tree_call_nonnegative_warnv_p): Added case for roundeven function.
        (integer_valued_real_call_p): Added case for roundeven function.
        * real.c (is_even): New function. Returns true if real number is even,
        otherwise returns false.
        (is_halfway_below): New function. Returns true if real number is
        halfway between two integers, else return false.
        (real_roundeven): New function. Round real number to nearest integer,
        rounding halfway cases towards even.
        * real.h (real_value): Added descriptive comments.  Added function
        declaration for roundeven function.
        * doc/extend.texi (Other Builtins): List roundeven variants among
        functions which can be handled as builtins.

gcc/testsuite/ChangeLog:

2019-08-26  Tejas Joshi  <tejasjoshi9673@gmail.com>

        * gcc.dg/torture/builtin-round-roundeven.c: New test.
        * gcc.dg/torture/builtin-round-roundevenf128.c: New test.

From-SVN: r274927

re PR target/91522 (STV is slow)

2019-08-26  Richard Biener  <rguenther@suse.de>

PR target/91522
PR target/91527
* config/i386/i386-features.h (general_scalar_chain::defs_map):
New member.
(general_scalar_chain::replace_with_subreg): Remove.
(general_scalar_chain::replace_with_subreg_in_insn): Likewise.
(general_scalar_chain::convert_reg): Adjust signature.
* config/i386/i386-features.c (scalar_chain::add_insn): Do not
iterate over all defs of a reg.
(general_scalar_chain::replace_with_subreg): Remove.
(general_scalar_chain::replace_with_subreg_in_insn): Likewise.
(general_scalar_chain::make_vector_copies): Populate defs_map,
place copy only after defs that are used as vectors in the chain.
(general_scalar_chain::convert_reg): Emit a copy for a specific
def in a specific instruction.
(general_scalar_chain::convert_op): All reg uses are converted here.
(general_scalar_chain::convert_insn): Emit copies for scalar
uses of defs here.  Replace uses with the copies we created.
Replace and convert the def.  Adjust REG_DEAD notes, remove
REG_EQUIV/EQUAL notes.
(general_scalar_chain::convert_registers): Only handle copies
into the chain here.

From-SVN: r274926

[PATCH 2/2] Add simplify rule for wrapped addition.

Add the transform (T)(A) + CST -> (T)(A + CST).  This enables vrp to
simplify sequences like

_2 = a_7 - 1;
_3 = (long unsigned int) _2;
_5 = _3 + 1

that ivopts creates.

--

gcc/ChangeLog:

2019-08-26  Robin Dapp  <rdapp@linux.ibm.com>

        * match.pd: Add (T)(A) + CST -> (T)(A + CST).

gcc/testsuite/ChangeLog:

2019-08-26  Robin Dapp  <rdapp@linux.ibm.com>

        * gcc.dg/tree-ssa/copy-headers-5.c: Do not run vrp pass.
        * gcc.dg/tree-ssa/copy-headers-7.c: Do not run vrp pass.
        * gcc.dg/tree-ssa/loop-15.c: Remove XFAIL.
        * gcc.dg/tree-ssa/pr23744.c: Change search pattern.
        * gcc.dg/wrapped-binop-simplify.c: New test.

From-SVN: r274925

[PATCH 1/2] Allow folding all statements.

This patch allows users of the substitute_and_fold_engine to
enable folding all statements.  It is now enabled for VRP which
is needed for the match.pd pattern in patch 2/2.

The loop versioning pass was missing one case when
deconstructing addresses that would only be triggered after
this patch for me:
It could handle addition and subsequent convert/nop but not
a convert/nop directly.  This would cause the hash to be
calculated differently and, in turn, cause the pass to miss
a versioning opportunity.  Fixed this by adding the missing
case.

--

gcc/ChangeLog:

2019-08-26  Robin Dapp  <rdapp@linux.ibm.com>

        * gimple-loop-versioning.cc (loop_versioning::record_address_fragment):
        Add nop_convert case.
        * tree-ssa-propagate.c (substitute_and_fold_dom_walker::before_dom_children):
        Fold all statements if requested.
        * tree-ssa-propagate.h (class substitute_and_fold_engine):
        Allow to fold all statements.
        * tree-vrp.c (class vrp_folder):
        Let substitute_and_fold_engine fold all statements.

From-SVN: r274923

re PR c/91526 (Unnecessary SSE and other instructions generated when compiling in C mode (vs. C++ mode))

2019-08-26  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91526
* passes.def: Note that after late FRE we do TODO_update_address_taken.
* tree-ssa-sccvn.c (pass_fre::execute): In late mode schedule
TODO_update_address_taken.

From-SVN: r274922

RISC-V: Add testcase for testing li pseudo instruction

gcc/testsuite/ChangeLog:

gcc.target/riscv/li.c: New test.

From-SVN: r274920

Daily bump.

From-SVN: r274919

gmm_malloc.h: Only use <errno.h> and errno if __STDC_HOSTED__.

* config/i386/gmm_malloc.h: Only use <errno.h> and errno if
__STDC_HOSTED__.

From-SVN: r274915

Daily bump.

From-SVN: r274907

[C++ PATCH] vfunc overrider simplification

https://gcc.gnu.org/ml/gcc-patches/2019-08/msg01674.html
cp/
* class.c (check_for_overrides): Conversion operators need
checking too.

testsuite/
* g++.dg/inherit/virtual14.C: New.

From-SVN: r274903

re PR fortran/91390 (treatment of extra parameter in a subroutine call)

2019-08-24  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/91390
PR fortran/91519
* frontend-passes.c (check_externals_procedure): New
function. If a procedure is not in the translation unit, create
an "interface" for it, including its formal arguments.
(check_externals_code): Use check_externals_procedure for common
code with check_externals_expr.
(check_externals_expr): Vice versa.
* gfortran.h (gfc_get_formal_from_actual-arglist): New prototype.
(gfc_compare_actual_formal): New prototype.
* interface.c (compare_actual_formal): Rename to
(gfc_compare_actual_formal): New function, make global.
(gfc_get_formal_from_actual_arglist): Make global, and move here from
* trans-types.c (get_formal_from_actual_arglist): Remove here.
(gfc_get_function_type): Use gfc_get_formal_from_actual_arglist.

2019-08-24  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/91390
PR fortran/91519
* gfortran.dg/bessel_3.f90: Add type mismatch errors.
* gfortran.dg/coarray_7.f90: Rename subroutines to avoid
additional errors.
* gfortran.dg/g77/20010519-1.f: Add -std=legacy. Remove
warnings for ASSIGN. Add warnings for type mismatch.
* gfortran.dg/goacc/acc_on_device-1.f95: Add -std=legacy.
Add catch-all warning.
* gfortran.dg/internal_pack_9.f90: Rename subroutine to
avoid type error.
* gfortran.dg/internal_pack_9.f90: Add -std=legacy. Add
warnings for type mismatch.
* gfortran.dg/pr39937.f: Add -std=legacy and type warnings. Move
here from
* gfortran.fortran-torture/compile/pr39937.f: Move to
gfortran.dg.

From-SVN: r274902

semantics.c (finish_switch_cond): Improve error message location.

/cp
2019-08-24  Paolo Carlini  <paolo.carlini@oracle.com>

* semantics.c (finish_switch_cond): Improve error message location.

/testsuite
2019-08-24  Paolo Carlini  <paolo.carlini@oracle.com>

* g++.dg/conversion/simd4.C: Test all the locations.

From-SVN: r274901

Daily bump.

From-SVN: r274898

Fix handling of namespace-scope undeduced auto decls.

* decl2.c (decl_dependent_p): New.
(mark_used): Check it instead of just processing_template_decl.

From-SVN: r274894

* parser.c (cp_parser_nested_name_specifier_opt): Avoid redundant error.

From-SVN: r274893

PR c++/91521 - wrong error with operator->.

* decl.c (grokdeclarator): Return error_mark_node for an invalid
trailing return type.

* g++.dg/parse/operator8.C: New test.

From-SVN: r274891

compiler: record pointer var values to remove write barriers

    Record when a local pointer variable is set to a value such that
    indirecting through the pointer does not require a write barrier.  Use
    that to eliminate write barriers when indirecting through that local
    pointer variable.  Only keep this information per-block, so it's not
    all that applicable.

    This reduces the number of write barriers generated when compiling the
    runtime package from 553 to 524.

    The point of this is to eliminate a bad write barrier in the bytes
    function in runtime/print.go.  Mark that function nowritebarrier so
    that the problem does not recur.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/191581

From-SVN: r274890

rs6000: New darn testcase (PR91481)

We used to implement darn with unspecs, not unspec_volatiles, which
means two darn instructions could be CSEd together.

This testcase tests it by adding together four random numbers.  If all
is well that means we get four darn instructions, because such a small
loop is unrolled fine at -O2 already.  If things go bad, combine will
combine it all to one darn and a shift left by two.

gcc/testsuite/
PR target/91481
* gcc.target/powerpc/darn-3.c: New testcase.

From-SVN: r274889

PR c++/79817 - attribute deprecated on namespace.

* cp-tree.h (cp_warn_deprecated_use_scopes): Declare.
* decl.c (grokdeclarator): Call cp_warn_deprecated_use_scopes.
(type_is_deprecated): Likewise.
* decl2.c (cp_warn_deprecated_use_scopes): New function.
* name-lookup.c (handle_namespace_attrs): Handle attribute deprecated.
* parser.c (cp_parser_namespace_alias_definition): Call
cp_warn_deprecated_use_scopes.
(cp_parser_using_declaration): Likewise.
(cp_parser_using_directive): Likewise.
* semantics.c (finish_id_expression_1): Likewise.

* g++.dg/cpp0x/attributes-namespace1.C: New test.
* g++.dg/cpp0x/attributes-namespace2.C: New test.
* g++.dg/cpp0x/attributes-namespace3.C: New test.
* g++.dg/cpp0x/attributes-namespace4.C: New test.
* g++.dg/cpp0x/attributes-namespace5.C: New test.
* g++.dg/cpp1z/namespace-attribs.C: Adjust.
* g++.dg/cpp1z/namespace-attribs2.C: Adjust.

From-SVN: r274888

Makefile.tpl (HOST_EXPORTS): Add CXX_FOR_BUILD.

* Makefile.tpl (HOST_EXPORTS): Add CXX_FOR_BUILD.
* Makefile.in: Regenerate.

From-SVN: r274887

Fix libstdc++ lstat missing return type for Windows target.

Building for i686-mingw32 target (with some local changes) produced an
error

error: ISO C++ forbids declaration of 'lstat' with no type [-fpermissive]

in libstdc++-v3/src/filesystem/ops-common.h.  This patch adds the
missing return type.  OK to commit (trunk and GCC 9 branch)?

Note 1: I haven't run the testsuite with this patch, but it fixes the
build failure I see (in sources with other local changes).

Note 2: I don't know why this hasn't produced build failures for other
people, but the missing return type looks wrong in any case even if
other local changes are for some reason needed for it to produce an
error.

* src/filesystem/ops-common.h [_GLIBCXX_FILESYSTEM_IS_WINDOWS]
(std::filesystem::__gnu_posix::lstat): Add return type.

From-SVN: r274885

* zh_CN.po: Update.

From-SVN: r274868

[C++ PATCH] vfunc overrider simplification

https://gcc.gnu.org/ml/gcc-patches/2019-08/msg01674.html
* class.c (check_for_override): Checking IDENTIFIER_VIRTUAL_P is
sufficient, reorder DECL_OVERRIDE_P check.

From-SVN: r274867

mips.md (mips_get_fcsr, [...]): Use SI machine mode for unspec_volatile operand.

        * config/mips/mips.md (mips_get_fcsr, *mips_get_fcsr): Use SI
        machine mode for unspec_volatile operand.

        * gcc.target/mips/get-fcsr-3.c: New test.

From-SVN: r274863

pr78973.c: Make functions static to avoid failures with -fpic

gcc/testsuite/ChangeLog:

* gcc.dg/pr78973.c: Make functions static to avoid failures
with -fpic
* gcc.dg/pr78973-2.c: Same.

From-SVN: r274860

Warray-bounds-36.c: Make functions static to avoid failures with -fpic.

gcc/testsuite/ChangeLog:

* gcc.dg/Warray-bounds-36.c: Make functions static to avoid failures
with -fpic.
* gcc.dg/Warray-bounds-41.c: Same.

From-SVN: r274859

[ARM] Deprecate -mneon-for-64bits

Deprecate -mneon-for-64bits since it no longer has any effect after
the DImode codegen improvements.

    gcc/
* gcc/doc/invoke.texi (mneon-for-64bits): Deprecate option.
* gcc/config/arm/arm.opt (mneon-for-64bits): Deprecate option.
* gcc/config/arm/arm.h (TARGET_PREFER_NEON_64BITS): Remove.
(prefer_neon_for_64bits): Remove.
* gcc/config/arm/arm.c (prefer_neon_for_64bits): Remove.
(tune_params): Remove PREF_NEON_64_FALSE uses.
(arm_option_override): Remove prefer_neon selection code.
(arm_print_tune_info): Remove prefer_neon_for_64bits.
* gcc/config/arm/arm-protos.h (tune_params): Remove
prefer_neon_for_64bits.
(prefer_neon_for_64bits): Remove.

From-SVN: r274858

PDF Layout

Prevent lines overlapping the right hand margin, this occurred for
references in the "See also:" section, it is prevented by placing
references on their own lines. Two example programs overlapped the
right hand margin: fixed by adjusting the position and of the comments.
Some tables had overlapping colums: the column widths were adjusted
accordingly.

From-SVN: r274857

[PATCH, c-family] Fix a PCH thinko (and thus PR61250).

When we are parsing a source file, the very first token might
be a PRAGMA_GCC_PCH_PREPROCESS.  This indicates that we are going
read in a PCH file (named as the value of the pragma).  If we don't
see this pragma, then we know that it's OK to release any resources
that the host might have set aside for the PCH file.

This fixes a thinko in the current implementation, in that the decision
to release resources was happening unconditionally right after the first
token is extracted but before it's been checked or acted upon.

This leads to the pch bug (seen on Darwin), because we actually do release
resources - which are subsequently (reasonably) assumed to be available
when reading a PCH file.  We then get random crashes or hangs depending
on the interaction between unmmap and malloc.

The bug is present everywhere but doesn't show on (say) Linux, since
the release of PCH resources is a NOP there.

This effects all the c-family front ends, because they all use c_lex_with_flags ()
to implement this.

The solution is to check for the PRAGMA_GCC_PCH_PREPROCESS and only call
c_common_no_more_pch () when that is not the first token.

A secondary effect of the collection is that the name of the PCH file
can be collected during the ggc_pch_read() reset of state.  Therefore
we should issue any diagnostic that might name the file before the
collections are triggered.

gcc/c-family/

2019-08-23  Iain Sandoe  <iain@sandoe.co.uk>

PR pch/61250
* c-lex.c (c_lex_with_flags):  Don't call
c_common_no_more_pch () from here.

gcc/c/

2019-08-23  Iain Sandoe  <iain@sandoe.co.uk>

PR pch/61250
* c-parser.c (c_parse_file): Call c_common_no_more_pch ()
after determining that the first token is not
PRAGMA_GCC_PCH_PREPROCESS.

gcc/cp/

2019-08-23  Iain Sandoe  <iain@sandoe.co.uk>

PR pch/61250
* parser.c (cp_parser_initial_pragma): Call c_common_no_more_pch ()
after determining that the first token is not
PRAGMA_GCC_PCH_PREPROCESS.

gcc/

2019-08-23  Iain Sandoe  <iain@sandoe.co.uk>

PR pch/61250
* ggc-page.c (ggc_pch_read): Read the ggc_pch_ondisk structure
and issue any diagnostics needed before collecting the pre-PCH
state.

From-SVN: r274856

re PR middle-end/91283 (gcc.dg/torture/c99-contract-1.c FAILs)

PR middle-end/91283
* common.opt (fexcess-precision=): Add Optimization flag.  Use
flag_excess_precision variable instead of
flag_excess_precision_cmdline.
* flags.h (class target_flag_state): Remove x_flag_excess_precision
member.
(flag_excess_precision): Don't define.
* langhooks.c (lhd_post_options): Set flag_excess_precision instead of
flag_excess_precision_cmdline.  Remove comment.
* opts.c (set_fast_math_flags): Use frontend_set_flag_excess_precision
and x_flag_excess_precision instead of
frontend_set_flag_excess_precision_cmdline and
x_flag_excess_precision_cmdline.
(fast_math_flags_set_p): Use x_flag_excess_precision instead of
x_flag_excess_precision_cmdline.
* toplev.c (init_excess_precision): Remove.
(lang_dependent_init_target): Don't call it.
ada/
* gcc-interface/misc.c (gnat_post_options): Set flag_excess_precision
instead of flag_excess_precision_cmdline.
brig/
* brig-lang.c (brig_langhook_post_options): Set flag_excess_precision
instead of flag_excess_precision_cmdline.
c-family/
* c-common.c (c_ts18661_flt_eval_method): Use flag_excess_precision
instead of flag_excess_precision_cmdline.
* c-cppbuiltin.c (c_cpp_flt_eval_method_iec_559): Likewise.
* c-opts.c (c_common_post_options): Likewise.
d/
* d-lang.cc (d_post_options): Set flag_excess_precision instead of
flag_excess_precision_cmdline.
fortran/
* options.c (gfc_post_options): Set flag_excess_precision instead of
flag_excess_precision_cmdline.  Remove comment.
go/
* go-lang.c (go_langhook_post_options): Set flag_excess_precision
instead of flag_excess_precision_cmdline.
lto/
* lto-lang.c (lto_post_options): Set flag_excess_precision instead of
flag_excess_precision_cmdline.  Remove comment.

From-SVN: r274850

Make -flto=jobserver parallel again.

2019-08-23  Martin Liska  <mliska@suse.cz>

* lto-wrapper.c (run_gcc): When setting jobserver
set also parallel to 1.  This was done so before r273908.

From-SVN: r274849

Intrinsic documentation layout changes.

Documentation of all intrinsics now have their sections in a consistent
order. Empty sections have been removed. Stray words have been removed.
Text in a wrong section has been moved.  A missing standard was inserted.

From-SVN: r274848

ABS Documentation fix

The specific intrinsics ZABS and CDABS return REAL(8) not COMPLEX(8).

From-SVN: r274847

re PR target/91306 ([MSP430] libgcc/crtstuff.c: Alignment of frame_dummy .init_array entry is too big)

2019-08-23  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

PR target/91306
* crtstuff.c (__CTOR_LIST__): Align to the "__alignof__" the array
element type, instead of "sizeof" the element type.
(__DTOR_LIST__): Likewise.
(__TMC_LIST__): Likewise.
(__do_global_dtors_aux_fini_array_entry): Likewise.
(__frame_dummy_init_array_entry): Likewise.
(__CTOR_END__): Likewise.
(__DTOR_END__): Likweise.
(__FRAME_END__): Likewise.
(__TMC_END__): Likewise.

From-SVN: r274846

[Arm] Add support for missing CPUs

This patch adds '-mcpu' options for following CPUs:
Cortex-M35P, Cortex-A77, Cortex-A76AE.

Related specifications are as following:
https://developer.arm.com/ip-products/processors/cortex-m
https://developer.arm.com/ip-products/processors/cortex-a

2019-08-23  Dennis Zhang  <dennis.zhang@arm.com>

* config/arm/arm-cpus.in (cortex-m35p): New entry.
(cortex-a76ae): Likewise.
(cortex-a77): Likewise
* config/arm/arm-tables.opt: Regenerate.
* config/arm/arm-tune.md: Likewise.
* doc/invoke.texi (ARM Options): Document cortex-m35p, cortx-a76ae,
cortex-a77 CPU options.

From-SVN: r274845

Clean up value-prof.c a bit.

2019-08-23  Martin Liska  <mliska@suse.cz>

* profile.c (instrument_values): Do not set
0 as last argument.
* tree-profile.c (gimple_gen_interval_profiler): Remove
last argument.
(gimple_gen_pow2_profiler): Likewise.
(gimple_gen_topn_values_profiler): Likewise.
(gimple_gen_ic_profiler): Likewise.
(gimple_gen_time_profiler): Likewise.
(gimple_gen_average_profiler): Likewise.
(gimple_gen_ior_profiler): Likewise.
* value-prof.c (dump_histogram_value): Use default
in switch statement instead of HIST_TYPE_MAX.
(stream_in_histogram_value): Likewise.
(gimple_duplicate_stmt_histograms): Do not
use NULL for implicitly set arguments.
(gimple_divmod_values_to_profile): Do not use
reserve+quick_push.
(gimple_indirect_call_to_profile): Likewise.
(gimple_find_values_to_profile): Use implicit
function call arguments.
* value-prof.h (gimple_alloc_histogram_value):
Set default values.
(gimple_gen_interval_profiler): Remove last argument.
(gimple_gen_pow2_profiler): Likewise.
(gimple_gen_topn_values_profiler): Likewise.
(gimple_gen_ic_profiler): Likewise.
(gimple_gen_time_profiler): Likewise.
(gimple_gen_average_profiler): Likewise.
(gimple_gen_ior_profiler): Likewise.

From-SVN: r274844

Daily bump.

From-SVN: r274843

PR c++/91304 - prefix attributes ignored in condition.

* parser.c (cp_parser_condition): Handle prefix attributes.

* g++.dg/cpp0x/gen-attrs-70.C: New test.

From-SVN: r274839

intrinsic.c (add_subroutines): ERRMSG is INTENT(INOUT) in co_broadcast, co_max, co_min, co_reduce, and co_sum.

2019-08-22  Steven G. Kargl  <kargl@gcc.gnu.org>

* intrinsic.c (add_subroutines): ERRMSG is INTENT(INOUT) in
co_broadcast, co_max, co_min, co_reduce, and  co_sum.

From-SVN: r274838

PR middle-end/91490 - bogus argument missing terminating nul warning on strlen of a flexible array member

gcc/c-family/ChangeLog:

PR middle-end/91490
* c-common.c (braced_list_to_string): Add argument and overload.
Handle flexible length arrays and unions.

gcc/testsuite/ChangeLog:

PR middle-end/91490
* c-c++-common/Warray-bounds-7.c: New test.
* gcc.dg/Warray-bounds-39.c: Expect either -Warray-bounds or
-Wstringop-overflow.
* gcc.dg/strlenopt-78.c: New test.

gcc/ChangeLog:

PR middle-end/91490
* builtins.c (c_strlen): Rename argument and introduce new local.
Set no-warning bit on original argument.
* expr.c (string_constant): Pass argument type to fold_ctor_reference.
Fold empty and zero constructors into empty strings.
* gimple-fold.c (fold_nonarray_ctor_reference): Return a STRING_CST
for missing initializers.
* tree.c (build_string_literal): Handle optional argument.
* tree.h (build_string_literal): Add defaulted argument.
* gimple-ssa-warn-restrict.c (maybe_diag_access_bounds): Check
no-warning bit on original expression.

From-SVN: r274837

Fix gcc.target/i386/minmax-4.c etc. on 32-bit Solaris/x86

* gcc.target/i386/minmax-4.c: Add -mno-stackrealign to dg-options.
* gcc.target/i386/minmax-5.c: Likewise.
* gcc.target/i386/minmax-6.c: Likewise.
* gcc.target/i386/minmax-7.c: Likewise.
* gcc.target/i386/pr91154.c: Likewise.

From-SVN: r274836

rs6000: Use unspec_volatile for darn (PR91481)

Every call to darn should deliver a *new* random number; such calls
should not be CSEd together.  So they should be unspec_volatile, not
plain unspec.

PR target/91481
* config/rs6000/rs6000.md (unspec): Delete UNSPEC_DARN, UNSPEC_DARN_32,
and UNSPEC_DARN_RAW.
(unspecv): New enumerator values UNSPECV_DARN, UNSPECV_DARN_32, and
UNSPECV_DARN_RAW.
(darn_32): Use an unspec_volatile, and UNSPECV_DARN_32.
(darn_raw): Use an unspec_volatile, and UNSPECV_DARN_RAW.
(darn): Use an unspec_volatile, and UNSPECV_DARN.

From-SVN: r274835

rs6000: Move various non-vector things out of altivec.md

* config/rs6000/altivec.md (unspec): Delete UNSPEC_DARN, UNSPEC_DARN_32,
UNSPEC_DARN_RAW, UNSPEC_CMPRB, UNSPEC_CMPRB2, UNSPEC_CMPEQB; move to...
* config/rs6000/rs6000.md (unspec): ... here.
* config/rs6000/altivec.md (darn_32, darn_raw, darn, cmprb,
*cmprb_internal, setb_signed, setb_unsigned, cmprb2, *cmprb2_internal,
cmpeqb, *cmpeqb_internal): Delete, move to...
* config/rs6000/rs6000.md (darn_32, darn_raw, darn, cmprb,
*cmprb_internal, setb_signed, setb_unsigned, cmprb2, *cmprb2_internal,
cmpeqb, *cmpeqb_internal): ... here.

From-SVN: r274834

c-parser.c (c_parser_declaration_or_fndef): Set DECL_ARGUMENTS of a FUNCTION_DECL to the right value in the presence of...

* c-parser.c (c_parser_declaration_or_fndef): Set DECL_ARGUMENTS of a
FUNCTION_DECL to the right value in the presence of nested declarators.

From-SVN: r274828

[arm] Fix use of CRC32 intrinsics with Armv8-a and hard-float

We currently have a nasty error when trying to use the __crc* intrinsics
with an -mfloat-abi=hard.
That is because the target pragma guarding them uses armv8-a+crc that
does not include fp by default.
So we get errors like:
error: '-mfloat-abi=hard': selected processor lacks an FPU

This patch fixes that by using an FP-enabled arch target pragma to guard
these intrinsics when floating-point is available.
That way both the softfloat and hardfloat variants work.

     * config/arm/arm_acle.h: Use arch=armv8-a+crc+simd pragma for CRC32
     intrinsics if __ARM_FP.
     Use __ARM_FEATURE_CRC32 ifdef guard.

     * gcc.target/arm/acle/crc_hf_1.c: New test.

From-SVN: r274827

[ARM] Remove remaining Neon DImode support

Remove the remaining Neon adddi3, subdi3 and negdi2 patterns.  As a result
adddi3, subdi3 and negdi2 can now always be expanded early irrespectively of
whether Neon is available.  Also expand the extenddi patterns at the same
time.  Several Neon arch attributes are no longer used and removed.

Code generation is improved in all cases, saving another 400-500 instructions
from the PR77308 testcase (total improvement is over 1700 instructions with
-mcpu=cortex-a57 -O2).

Bootstrap & regress OK on arm-none-linux-gnueabihf --with-cpu=cortex-a57

    gcc/
* config/arm/arm.md (neon_for_64bits): Remove.
(avoid_neon_for_64bits): Remove.
(arm_adddi3): Always split early.
(arm_subdi3): Always split early.
(negdi2): Remove Neon expansion.
(split zero_extend): Split before reload.
(split sign_extend): Split before reload.

From-SVN: r274825

[ARM] Cleanup DImode shifts

Like the logical operations, expand all shifts early rather than only
sometimes.  The Neon shift expansions are never emitted (not even with
-fneon-for-64bits), so they are not useful.  So all the late expansions
and Neon shift patterns can be removed, and shifts are more optimized
as a result.  Since some extend patterns use Neon DImode shifts, remove
the Neon extend variants and related splits.

A simple example now generates the same efficient code after this
patch with -mfpu=neon and -mfpu=vfp (previously just the fact of
having Neon enabled resulted inefficient code for no reason).

unsigned long long f(unsigned long long x, unsigned long long y)
{ return x & (y >> 33); }

Before:
strd    r4, r5, [sp, #-8]!
lsr     r4, r3, #1
mov     r5, #0
and     r1, r1, r5
and     r0, r0, r4
ldrd    r4, r5, [sp]
add     sp, sp, #8
bx      lr

After:
and     r0, r0, r3, lsr #1
mov     r1, #0
bx      lr

Bootstrap and regress OK on arm-none-linux-gnueabihf --with-cpu=cortex-a57

    gcc/
* config/arm/iterators.md (qhs_extenddi_cstr): Update.
(qhs_extenddi_cstr): Likewise.
* config/arm/arm.md (ashldi3): Always expand early.
(ashlsi3): Likewise.
(ashrsi3): Likewise.
(zero_extend<mode>di2): Remove Neon variants.
(extend<mode>di2): Likewise.
* config/arm/neon.md (ashldi3_neon_noclobber): Remove.
(signed_shift_di3_neon): Likewise.
(unsigned_shift_di3_neon): Likewise.
(ashrdi3_neon_imm_noclobber): Likewise.
(lshrdi3_neon_imm_noclobber): Likewise.
(<shift>di3_neon): Likewise.
(split extend): Remove DI extend split patterns.

   gcc/testsuite/
* gcc.target/arm/neon-extend-1.c: Remove test.
* gcc.target/arm/neon-extend-2.c: Remove test.

From-SVN: r274824

[ARM] Cleanup logical DImode operations

Cleanup the logical DImode operations since the current implementation is way
too complicated.  Thumb-1, Thumb-2, VFP/Neon and iwMMXt all work differently,
resulting in a bewildering number of expansions, patterns and splits across
several md files.  All this complexity is counterproductive and results in
inefficient code.

A much simpler approach is to split these operations early in the expander
so that optimizations and register allocation are applied on the 32-bit halves.
Codegeneration is unchanged on Thumb-1 and Arm/Thumb-2 without Neon or iwMMXt
(which already expand these instructions early).  With Neon these changes save
~1000 instructions from the PR77308 testcase, mostly by significantly reducing
register pressure and spilling.

Bootstrap OK on arm-none-linux-gnueabihf --with-cpu=cortex-a57

    gcc/
* config/arm/arm.md (split and/eor/ior): Remove Neon check.
(split not): Add DImode not splitter.
(anddi3): Remove pattern.
(anddi3_insn): Likewise.
(anddi_zesidi_di): Likewise.
(anddi_sesdi_di): Likewise.
(anddi_notdi_di): Likewise.
(anddi_notzesidi_di): Likewise.
(anddi_notsesidi_di): Likewise.
(iordi3): Likewise.
(iordi3_insn): Likewise.
(iordi_zesidi_di): Likewise.
(iordi_sesidi_di): Likewise.
(xordi3): Likewise.
(xordi3_insn): Likewise.
(xordi_sesidi_di): Likewise.
(xordi_zesidi_di): Likewise.
(one_cmpldi2): Likewise.
(one_cmpldi2_insn): Likewise.
* config/arm/constraints.md: Remove De, Df, Dg constraints.
* config/arm/iwmmxt.md (iwmmxt_iordi3): Remove general register
alternative.
(iwmmxt_xordi3): Likewise.
(iwmmxt_anddi3): Likewise.
* config/arm/neon.md (orndi3_neon): Remove pattern.
(anddi_notdi_di): Likewise.
* config/arm/predicates.md (arm_anddi_operand_neon): Remove.
(arm_iordi_operand_neon): Likewise.
(arm_xordi_operand_neon): Likewise.
* config/arm/thumb2.md(iordi_notdi_di): Remove pattern.
(iordi_notzesidi_di): Likewise.
(iordi_notdi_zesidi): Likewise.
(iordi_notsesidi_di): Likewise.

From-SVN: r274823

[Arm] Add 16-bit thumb alternatives to iorsi3_compare0[_scratch]

The iorsi3_compare0 and iorsi3_compare0_scratch patterns can make use
of the 16-bit thumb orrs instruction if suitable registers are
allocated.  This patch adds the alternative to allow this to happen.

* config/arm/arm.md (iorsi3_compare0): Add alternative for 16-bit thumb
insn.
(iorsi3_compare0_scratch): Likewise.

From-SVN: r274822

[aarch64]: add intrinsics for vld1(q)_x4 and vst1(q)_x4

This patch adds the intrinsic functions for:
- vld1_<mode>_x4
- vst1_<mode>_x4
- vld1q_<mode>_x4
- vst1q_<mode>_x4

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

Committed on behalf of Sylvia Taylor.

2019-08-22  Sylvia Taylor  <sylvia.taylor@arm.com>

* config/aarch64/aarch64-simd-builtins.def:
(ld1x4): New.
(st1x4): Likewise.
* config/aarch64/aarch64-simd.md:
(aarch64_ld1x4<VALLDIF:mode>): New pattern.
(aarch64_st1x4<VALLDIF:mode>): Likewise.
(aarch64_ld1_x4_<mode>): Likewise.
(aarch64_st1_x4_<mode>): Likewise.
* config/aarch64/arm_neon.h:
(vld1_s8_x4): New function.
(vld1q_s8_x4): Likewise.
(vld1_s16_x4): Likewise.
(vld1q_s16_x4): Likewise.
(vld1_s32_x4): Likewise.
(vld1q_s32_x4): Likewise.
(vld1_u8_x4): Likewise.
(vld1q_u8_x4): Likewise.
(vld1_u16_x4): Likewise.
(vld1q_u16_x4): Likewise.
(vld1_u32_x4): Likewise.
(vld1q_u32_x4): Likewise.
(vld1_f16_x4): Likewise.
(vld1q_f16_x4): Likewise.
(vld1_f32_x4): Likewise.
(vld1q_f32_x4): Likewise.
(vld1_p8_x4): Likewise.
(vld1q_p8_x4): Likewise.
(vld1_p16_x4): Likewise.
(vld1q_p16_x4): Likewise.
(vld1_s64_x4): Likewise.
(vld1_u64_x4): Likewise.
(vld1_p64_x4): Likewise.
(vld1q_s64_x4): Likewise.
(vld1q_u64_x4): Likewise.
(vld1q_p64_x4): Likewise.
(vld1_f64_x4): Likewise.
(vld1q_f64_x4): Likewise.
(vst1_s8_x4): Likewise.
(vst1q_s8_x4): Likewise.
(vst1_s16_x4): Likewise.
(vst1q_s16_x4): Likewise.
(vst1_s32_x4): Likewise.
(vst1q_s32_x4): Likewise.
(vst1_u8_x4): Likewise.
(vst1q_u8_x4): Likewise.
(vst1_u16_x4): Likewise.
(vst1q_u16_x4): Likewise.
(vst1_u32_x4): Likewise.
(vst1q_u32_x4): Likewise.
(vst1_f16_x4): Likewise.
(vst1q_f16_x4): Likewise.
(vst1_f32_x4): Likewise.
(vst1q_f32_x4): Likewise.
(vst1_p8_x4): Likewise.
(vst1q_p8_x4): Likewise.
(vst1_p16_x4): Likewise.
(vst1q_p16_x4): Likewise.
(vst1_s64_x4): Likewise.
(vst1_u64_x4): Likewise.
(vst1_p64_x4): Likewise.
(vst1q_s64_x4): Likewise.
(vst1q_u64_x4): Likewise.
(vst1q_p64_x4): Likewise.
(vst1_f64_x4): Likewise.
(vst1q_f64_x4): Likewise.

* gcc.target/aarch64/advsimd-intrinsics/vld1x4.c: New test.
* gcc.target/aarch64/advsimd-intrinsics/vst1x4.c: New test.

From-SVN: r274820

aarch64-sve.md (vcond_mask): Add "@".

2019-08-22  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>

* config/aarch64/aarch64-sve.md (vcond_mask): Add "@".

From-SVN: r274817

Daily bump.

From-SVN: r274816

re PR target/88839 ([SVE] Poor implementation of blend-like permutes)

2019-08-22  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>
    Richard Sandiford  <richard.sandiford@arm.com>

PR target/88839
* config/aarch64/aarch64.c (aarch64_evpc_sel): New function.
(aarch64_expand_vec_perm_const_1): Call aarch64_evpc_sel.

testsuite/
* gcc.target/aarch64/sve/sel_1.c: New test.
* gcc.target/aarch64/sve/sel_2.c: Likewise.
* gcc.target/aarch64/sve/sel_3.c: Likewise.
* gcc.target/aarch64/sve/sel_4.c: Likewise.
* gcc.target/aarch64/sve/sel_5.c: Likewise.
* gcc.target/aarch64/sve/sel_6.c: Likewise.

Co-Authored-By: Richard Sandiford <richard.sandiford@arm.com>
From-SVN: r274810

re PR target/90724 (ICE with __sync_bool_compare_and_swap with -march=armv8.2-a+sve)

2019-08-21  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>

PR target/90724
* config/aarch64/aarch64.c (aarch64_gen_compare_reg_maybe_ze): Force y
in reg if it fails aarch64_plus_operand predicate.

From-SVN: r274805

compiler: don't use pkgpath for fieldtrack of unexported field

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/191040

From-SVN: r274803

compiler: if hidden function referenced by inline, don't hide descriptor

    Fixes golang/go#33739

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/191037

From-SVN: r274800

[C++] Protect call to copy_attributes_to_builtin (PR91505)

copy_attributes_to_builtin only handles BUILT_IN_NORMAL, but C++ was
calling it immediately after the:

  if (DECL_BUILT_IN_CLASS (newdecl) == BUILT_IN_NORMAL)

block.  The corresponding C code calls it inside the block instead.

2019-08-21  Richard Sandiford  <richard.sandiford@arm.com>

gcc/cp/
PR c++/91505
* decl.c (duplicate_decls): Call copy_attributes_to_builtin inside
the BUILT_IN_NORMAL block rather than afterward.

gcc/testsuite/
PR c++/91505
* g++.target/i386/crc32-4.C: New test.

From-SVN: r274799

re PR tree-optimization/91482 (__builtin_assume_aligned should not break write combining)

2019-08-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91482
* tree-ssa-ccp.c (ccp_folder::fold_stmt): Remove useless
BUILT_IN_ASSUME_ALIGNED calls.

* gcc.dg/tree-ssa/pr91482.c: New testcase.

From-SVN: r274796

c-ada-spec.c (dump_ada_function_declaration): Be prepared for broken function declarations where arguments are missing.

* c-ada-spec.c (dump_ada_function_declaration): Be prepared for broken
function declarations where arguments are missing.  Rename variables.

From-SVN: r274794