[riscv] Wrap ASM_OUTPUT_LABELREF in do {} while (0)

2017-11-12  Tom de Vries  <tom@codesourcery.com>

* config/riscv/riscv.h (ASM_OUTPUT_LABELREF): Wrap in do {} while (0).

From-SVN: r254666

Remove semicolon after ASM_OUTPUT_ASCII

2017-11-12  Tom de Vries  <tom@codesourcery.com>

* config/elfos.h (ASM_OUTPUT_ASCII): Remove semicolon after macro body.

From-SVN: r254665

[cr16, powerpcspe, rs6000] Remove semicolon after ASM_OUTPUT_LABELREF macro body

2017-11-12  Tom de Vries  <tom@codesourcery.com>

* config/cr16/cr16.h (ASM_OUTPUT_LABELREF): Remove semicolon after macro
body.
* config/powerpcspe/xcoff.h (ASM_OUTPUT_LABELREF): Same.
* config/rs6000/xcoff.h (ASM_OUTPUT_LABELREF): Same.
* defaults.h (ASM_OUTPUT_LABELREF): Same.

From-SVN: r254664

Daily bump.

From-SVN: r254663

re PR fortran/82932 ([OOP] ICE in update_compcall_arglist, at fortran/resolve.c:5837)

2017-11-11  Janus Weil  <janus@gcc.gnu.org>

PR fortran/82932
* resolve.c (update_compcall_arglist): Improve error recovery,
remove a gcc_assert.

2017-11-11  Janus Weil  <janus@gcc.gnu.org>

PR fortran/82932
* gfortran.dg/typebound_call_29.f90: New test.

From-SVN: r254660

re PR c/81117 (Improve buffer overflow checking in strncpy)

gcc/ChangeLog:

PR c/81117
* doc/extend.texi (attribute nonstring): Remove spurious argument.

From-SVN: r254659

PR bootstrap/82948 - prefix.c:202:15: error: 'char* strncpy(char*, const char*,

   size_t)' destination unchanged after copying no bytes

gcc/ChangeLog:

PR bootstrap/82948
* prefic.c (translate_name): Replace strncpy with memcpy to
avoid -Wstringop-truncation.

From-SVN: r254658

Add forgotten tests.

From-SVN: r254657

* tree-ssa-loop-im.c (execute_sm_if_changed): Do not compute freq_sum.

From-SVN: r254653

* predict.c (maybe_hot_frequency_p): Do not use cfun.

From-SVN: r254652

* predict.c (maybe_hot_frequency_p): Do not use cfun.

From-SVN: r254651

tree-ssa-tail-merge.c (replace_block_by): Fix and re-enable profile merging.

* tree-ssa-tail-merge.c (replace_block_by): Fix and re-enable profile
merging.

From-SVN: r254650

Remove ada/ prefix

From-SVN: r254649

re PR fortran/82886 (ICE with -finit-derived in gfc_conv_expr, at fortran/trans-expr.c:7807)

2017-11-10  Fritz Reese <fritzoreese@gmail.com>

    PR fortran/82886

    gcc/fortran/ChangeLog:

PR fortran/82886
* gfortran.h (gfc_build_init_expr): New prototype.
* invoke.texi (finit-derived): Update documentation.
* expr.c (gfc_build_init_expr): New, from gfc_build_default_init_expr.
(gfc_build_default_init_expr): Redirect to gfc_build_init_expr(,,false)
(component_initializer): Force building initializers using
gfc_build_init_expr(,,true).

    gcc/testsuite/ChangeLog:

PR fortran/82886
* gfortran.dg/init_flag_16.f03: New testcase.

From-SVN: r254648

Daily bump.

From-SVN: r254647

Minor tweak to libstdc++ testsuite helper type

* testsuite/util/testsuite_tr1.h (ThrowMoveConsClass): Use noexcept.

From-SVN: r254644

rs6000.md (bswaphi2_reg): On ISA 3.0 systems, enable generating XXBRH if the value is in a vector register.

[gcc]
2017-11-10  Michael Meissner  <meissner@linux.vnet.ibm.com>

* config/rs6000/rs6000.md (bswaphi2_reg): On ISA 3.0 systems,
enable generating XXBRH if the value is in a vector register.
(bswapsi2_reg): On ISA 3.0 systems, enable generating XXBRW if the
value is in a vector register.
(bswapdi2_reg): On ISA 3.0 systems, always use XXBRD to do
register to register bswap64's instead of doing the GPR sequence
used on previous machines.
(bswapdi2_xxbrd): New insn.
(bswapdi2_reg): Disallow on ISA 3.0.
(register to register bswap64 splitter): Do not split the insn on
ISA 3.0 systems that use XXBRD.

[gcc/testsuite]
2017-11-10  Michael Meissner  <meissner@linux.vnet.ibm.com>

* gcc.target/powerpc/p9-xxbr-3.c: New test.

From-SVN: r254643

Fix my last ChangeLog entry

From-SVN: r254642

re PR c/81117 (Improve buffer overflow checking in strncpy)

gcc/ChangeLog:

PR c/81117
* config/darwin-c.c (framework_construct_pathname): Replace strncpy
with memcpy.
(find_subframework_file): Same.

From-SVN: r254641

Add plugin API for processing plugin-added input files.

Gold plugins may wish to further process an input file added by a plugin. For
example, the plugin may need to assign a unique segment for sections in a
plugin-generated input file. This patch adds a plugin callback that the linker
will call when reading symbols from a new input file added after the
all_symbols_read event (i.e. an input file added by a plugin).

2017-11-10  Stephen Crane <sjc@immunant.com>

* plugin-api.h: Add plugin API for processing plugin-added
input files.

From-SVN: r254640

vr-values.h (VR_INITIALIZER): Move #define here.

* vr-values.h (VR_INITIALIZER): Move #define here.
* gimple-ssa-evrp.c: New file with contents extracted from tree-vrp.c
* Makefile.in (OBJS): Add tree-evrp.o
* tree-vrp.h (assert_info): Move structure definition here.
(set_value_range_to_varying): Prototype.
(vrp_operand_equal_p, range_includes_zero_p): Likewise.
(infer_value_range, register_edge_assert_for): Likewise.
(stmt_interesting_for_vrp): Likewise.
* tree-vrp.c: Move all methods for evrp class into tree-evrp.c.
(set_value_range_to_varying): No longer static.
(vrp_operand_equal_p, range_includes_zero_p): Likewise.
(infer_value_range, register_edge_assert_for): Likewise.

From-SVN: r254639

PR libstdc++/82917 add missing returns in <fstream>

PR libstdc++/82917
* include/std/fstream (basic_ifstream::open, basic_ofstream::open)
(basic_fstream::open): Fix missing return.

From-SVN: r254638

auto-profile.c (afdo_indirect_call): Drop frequency.

* auto-profile.c (afdo_indirect_call): Drop frequency.
* cgraph.c (symbol_table::create_edge): Drop frequency argument.
(cgraph_node::create_edge): Drop frequency argument.
(cgraph_node::create_indirect_edge): Drop frequency argument.
(cgraph_edge::make_speculative): Drop frequency arguments.
(cgraph_edge::resolve_speculation): Do not update frequencies
(cgraph_edge::dump_edge_flags): Do not dump frequency.
(cgraph_node::dump): Check consistency in IPA mode.
(cgraph_edge::maybe_hot_p): Use IPA counter.
(cgraph_edge::verify_count_and_frequency): Rename to ...
(cgraph_edge::verify_count): ... this one; drop frequency checking.
(cgraph_node::verify_node): Update.
* cgraph.h (struct cgraph_edge): Drop frequency.
(cgraph_edge::frequency): New function.
* cgraphbuild.c (pass_build_cgraph_edges::execute): Donot pass
frequencies.
(cgraph_edge::rebuild_edges): Likewise.
* cgraphclones.c (cgraph_edge::clone): Scale only counts.
(duplicate_thunk_for_node): Do not pass frequency.
(cgraph_node::create_clone): Scale only counts.
(cgraph_node::create_virtual_clone): Do not pass frequency.
(cgraph_node::create_edge_including_clones): Do not pass frequency.
(cgraph_node::create_version_clone): Do not pass frequency.
* cgraphunit.c (cgraph_node::analyze): Do not pass frequency.
(cgraph_node::expand_thunk): Do not pass frequency.
(cgraph_node::create_wrapper): Do not pass frequency.
* gimple-iterator.c (update_call_edge_frequencies): Do not pass
frequency.
* gimple-streamer-in.c (input_bb): Scale only IPA counts.
* ipa-chkp.c (chkp_produce_thunks): Do not pass frequency.
* ipa-cp.c (ipcp_lattice::print): Use frequency function.
(gather_caller_stats): Use frequency function.
(ipcp_cloning_candidate_p): Use frequency function.
(ipcp_propagate_stage): Use frequency function.
(get_info_about_necessary_edges): Use frequency function.
(update_profiling_info): Update only IPA profile.
(update_specialized_profile): Use frequency functoin.
(perhaps_add_new_callers): Update only IPA profile.
* ipa-devirt.c (ipa_devirt): Use IPA profile.
* ipa-fnsummary.c (redirect_to_unreachable): Do not set frequrency.
(dump_ipa_call_summary): Use frequency function.
(estimate_edge_size_and_time): Use frequency function.
(ipa_merge_fn_summary_after_inlining): Use frequency function.
* ipa-inline-analysis.c (do_estimate_edge_time): Use IPA profile.
* ipa-inline-transform.c (update_noncloned_frequencies): Rename to ..
(update_noncloned_counts): ... ths one; scale counts only.
(clone_inlined_nodes): Do not scale frequency.
(inline_call): Do not pass frequency.
* ipa-inline.c (compute_uninlined_call_time): Use IPA profile.
(compute_inlined_call_time): Use IPA profile.
(want_inline_small_function_p): Use IPA profile.
(want_inline_self_recursive_call_p): Use IPA profile.
(edge_badness): Use IPA profile.
(lookup_recursive_calls): Use IPA profile.
(recursive_inlining): Do not pass frequency.
(resolve_noninline_speculation): Do not update frequency.
(inline_small_functions): Collect max of IPA profile.
(dump_overall_stats): Dump IPA porfile.
(dump_inline_stats): Dump IPA porfile.
(ipa_inline): Collect IPA stats.
* ipa-inline.h (clone_inlined_nodes): Update prototype.
* ipa-profile.c (ipa_propagate_frequency_1): Use frequency function.
(ipa_propagate_frequency): Use frequency function.
(ipa_profile): Cleanup.
* ipa-prop.c (ipa_make_edge_direct_to_target): Do not pass frequency
* ipa-utils.c (ipa_merge_profiles): Merge all profiles.
* lto-cgraph.c (lto_output_edge): Do not stream frequency.
(input_node): Do not stream frequency.
(input_edge): Do not stream frequency.
(merge_profile_summaries): Scale only IPA profiles.
* omp-simd-clone.c (simd_clone_adjust): Do not pass frequency.
* predict.c (drop_profile): Do not recompute frequency.
* trans-mem.c (ipa_tm_insert_irr_call): Do not pass frequency.
(ipa_tm_insert_gettmclone_call): Do not pass frequency.
* tree-cfg.c (execute_fixup_cfg): Drop profile to global0 if needed.
* tree-chkp.c (chkp_copy_bounds_for_assign): Do not pass frequency.
* tree-emutls.c (gen_emutls_addr): Do not pass frequency.
* tree-inline.c (copy_bb): Do not scale frequency.
(expand_call_inline): Do not scale frequency.
(tree_function_versioning): Do not scale frequency.
* ubsan.c (ubsan_create_edge): Do not pass frequency.

lto/ChangeLog:

2017-11-10  Jan Hubicka  <hubicka@ucw.cz>

* lto-partition.c (lto_balanced_map): Use frequency accessor.

From-SVN: r254636

force-indirect-call-1.c: Merge scan strings.

* gcc.target/i386/force-indirect-call-1.c: Merge scan strings.
* gcc.target/i386/force-indirect-call-2.c: Ditto.
Require fpic effective target.
* gcc.target/i386/force-indirect-call-3.c: Ditto.
Require lp64 effective target.

From-SVN: r254635

GFNI enabling [3/4]

gcc/
 * config/i386/gfniintrin.h (_mm_gf2p8affine_epi64_epi8,
 _mm256_gf2p8affine_epi64_epi8, _mm_mask_gf2p8affine_epi64_epi8,
 _mm_maskz_gf2p8affine_epi64_epi8, _mm256_mask_gf2p8affine_epi64_epi8,
 _mm256_maskz_gf2p8affine_epi64_epi8,
 _mm512_mask_gf2p8affine_epi64_epi8, _mm512_gf2p8affine_epi64_epi8
 _mm512_maskz_gf2p8affine_epi64_epi8): New intrinsics.
 * config/i386/i386-builtin.def (__builtin_ia32_vgf2p8affineqb_v64qi,
 __builtin_ia32_vgf2p8affineqb_v32qi,
 __builtin_ia32_vgf2p8affineqb_v16qi): New builtins.
 * config/i386/sse.md (vgf2p8affineqb_<mode><mask_name>): New pattern.
gcc/testsuite/
 * gcc.target/i386/avx-1.c: Handle new intrinsics.
 * gcc.target/i386/avx512f-gf2p8affineqb-2.c: New runtime tests.
 * gcc.target/i386/avx512vl-gf2p8affineqb-2.c: Ditto.
 * gcc.target/i386/gfni-1.c: Add tests for GF2P8AFFINE.
 * gcc.target/i386/gfni-2.c: Ditto.
 * gcc.target/i386/gfni-3.c: Ditto.
 * gcc.target/i386/gfni-4.c: Ditto.
 * gcc.target/i386/sse-13.c: Handle new tests.
 * gcc.target/i386/sse-14.c: Handle new tests.
 * gcc.target/i386/sse-23.c: Handle new tests.

From-SVN: r254634

[testsuite/ARM] Consolidate sources for cmse tests

For the most part, testcases under gcc.target/arm/cmse/baseline and
gcc.target/arm/cmse/mainline are duplicate copies with only different
dejagnu directives. Although there is no requirement for them to be
similar, having them both identical allow to compare the code generated
and make it easier in case of change in code generation to both
architecture to update the testcases (if one needs updating so does the
other).

Similarly all the tests in gcc.target/arm/cmse/mainline/<floatabi> have
the same source but are duplicate copies.

This patch moves all the code in the tests to a parent directory:
gcc.target/arm/cmse for tests shared by Armv8-M Baseline and Mainline
and gcc.target/arm/cmse/mainline for tests *only* shared by the various
float ABI of Armv8-M Mainline. C includes are then used where the code
used to sit.

Note that the cmse-13.c test used to differ slightly between
architectures and float ABI tested in the first floating-point constant
passed to bar: sometimes 1.0 and sometimes 3.0. This patch settles on
3.0 to not confuse with the 1.0 constant used to clear VFP registers in
some of the configurations.

2017-11-10  Thomas Preud'homme  <thomas.preudhomme@arm.com>

    gcc/testsuite/
    * gcc.target/arm/cmse/bitfield-4.x: New file.
    * gcc.target/arm/cmse/baseline/bitfield-4.c: Remove code and include
    above file.
    * gcc.target/arm/cmse/mainline/bitfield-4.c: Likewise.
    * gcc.target/arm/cmse/bitfield-5.x: New file.
    * gcc.target/arm/cmse/baseline/bitfield-5.c: Remove code and include
    above file.
    * gcc.target/arm/cmse/mainline/bitfield-5.c: Likewise.
    * gcc.target/arm/cmse/bitfield-6.x: New file.
    * gcc.target/arm/cmse/baseline/bitfield-6.c: Remove code and include
    above file.
    * gcc.target/arm/cmse/mainline/bitfield-6.c: Likewise.
    * gcc.target/arm/cmse/bitfield-7.x: New file.
    * gcc.target/arm/cmse/baseline/bitfield-7.c: Remove code and include
    above file.
    * gcc.target/arm/cmse/mainline/bitfield-7.c: Likewise.
    * gcc.target/arm/cmse/bitfield-8.x: New file.
    * gcc.target/arm/cmse/baseline/bitfield-8.c: Remove code and include
    above file.
    * gcc.target/arm/cmse/mainline/bitfield-8.c: Likewise.
    * gcc.target/arm/cmse/bitfield-9.x: New file.
    * gcc.target/arm/cmse/baseline/bitfield-9.c: Remove code and include
    above file.
    * gcc.target/arm/cmse/mainline/bitfield-9.c: Likewise.
    * gcc.target/arm/cmse/bitfield-and-union.x: New file.
    * gcc.target/arm/cmse/baseline/bitfield-and-union-1.c: Rename into ...
    * gcc.target/arm/cmse/baseline/bitfield-and-union.c: This.  Remove code
    and include above bitfield-and-union.x file.
    * gcc.target/arm/cmse/mainline/bitfield-and-union-1.c: Rename into ...
    * gcc.target/arm/cmse/mainline/bitfield-and-union.c: this.  Remove code
    and include above bitfield-and-union.x file.
    * gcc.target/arm/cmse/cmse-13.x: New file.
    * gcc.target/arm/cmse/baseline/cmse-13.c: Remove code and include above
    file.
    * gcc.target/arm/cmse/mainline/hard-sp/cmse-13.c: Likewise.
    * gcc.target/arm/cmse/mainline/hard/cmse-13.c: Likewise.
    * gcc.target/arm/cmse/mainline/soft/cmse-13.c: Likewise.
    * gcc.target/arm/cmse/mainline/softfp/cmse-13.c: Likewise.
    * gcc.target/arm/cmse/cmse-5.x: New file.
    * gcc.target/arm/cmse/mainline/hard-sp/cmse-5.c: Remove code and
    include above file.
    * gcc.target/arm/cmse/mainline/hard/cmse-5.c: Likewise.
    * gcc.target/arm/cmse/mainline/soft/cmse-5.c: Likewise.
    * gcc.target/arm/cmse/mainline/softfp-sp/cmse-5.c: Likewise.
    * gcc.target/arm/cmse/mainline/softfp/cmse-5.c: Likewise.
    * gcc.target/arm/cmse/cmse-7.x: New file.
    * gcc.target/arm/cmse/mainline/hard-sp/cmse-7.c: Remove code and
    include above file.
    * gcc.target/arm/cmse/mainline/hard/cmse-7.c: Likewise.
    * gcc.target/arm/cmse/mainline/soft/cmse-7.c: Likewise.
    * gcc.target/arm/cmse/mainline/softfp-sp/cmse-7.c: Likewise.
    * gcc.target/arm/cmse/mainline/softfp/cmse-7.c: Likewise.
    * gcc.target/arm/cmse/cmse-8.x: New file.
    * gcc.target/arm/cmse/mainline/hard-sp/cmse-8.c: Remove code and
    include above file.
    * gcc.target/arm/cmse/mainline/hard/cmse-8.c: Likewise.
    * gcc.target/arm/cmse/mainline/soft/cmse-8.c: Likewise.
    * gcc.target/arm/cmse/mainline/softfp-sp/cmse-8.c: Likewise.
    * gcc.target/arm/cmse/mainline/softfp/cmse-8.c: Likewise.
    * gcc.target/arm/cmse/union-1.x: New file.
    * gcc.target/arm/cmse/baseline/union-1.c: Remove code and include above
    file.
    * gcc.target/arm/cmse/mainline/union-1.c: Likewise.
    * gcc.target/arm/cmse/union-2.x: New file.
    * gcc.target/arm/cmse/baseline/union-2.c: Remove code and include above
    file.
    * gcc.target/arm/cmse/mainline/union-2.c: Likewise.

From-SVN: r254633

re PR target/82641 (Unable to enable crc32 for a certain function with target attribute on ARM (aarch32))

2017-11-10  Tamar Christina  <tamar.christina@arm.com>

PR target/82641
* config/arm/arm.c
(arm_option_override): Refactor.
(arm_option_reconfigure_globals): New.
(arm_options_perform_arch_sanity_checks): New.
* config/arm/arm-protos.h (arm_option_reconfigure_globals): New prototype.
(arm_options_perform_arch_sanity_checks): Likewise

From-SVN: r254632

power9.md (power9-qpdiv): Correct DFU pipe usage.

* rs6000/power9.md (power9-qpdiv): Correct DFU pipe usage.
(power9-qpmul): New.
* rs6000/rs6000.md ("type" attr): Add qmul.
(mul<mode>3, fma<mode>4_hw, *fms<mode>4_hw, *nfma<mode>4_hw,
*nfms<mode>4_hw, mul<mode>3_odd, fma<mode>4_odd, *fms<mode>4_odd,
*nfma<mode>4_odd, *nfms<mode>4_odd): Change type to qmul.

From-SVN: r254631

PR c/81117 - Improve buffer overflow checking in strncpy

gcc/ChangeLog:

PR c/81117
* builtins.c (compute_objsize): Handle arrays that
compute_builtin_object_size likes to fail for.  Make extern.
* builtins.h (compute_objsize): Declare.
(check_strncpy_sizes): New function.
(expand_builtin_strncpy): Call check_strncpy_sizes.
* gimple-fold.c (gimple_fold_builtin_strncpy): Implement
-Wstringop-truncation.
(gimple_fold_builtin_strncat): Same.
* gimple.c (gimple_build_call_from_tree): Set call location.
* tree-ssa-strlen.c (strlen_to_stridx): New global variable.
(maybe_diag_bound_equal_length, is_strlen_related_p): New functions.
(handle_builtin_stxncpy, handle_builtin_strncat): Same.
(handle_builtin_strlen): Use strlen_to_stridx.
(strlen_optimize_stmt): Handle flavors of strncat, strncpy, and
stpncpy.
Use strlen_to_stridx.
(pass_strlen::execute): Release strlen_to_stridx.
* doc/invoke.texi (-Wsizeof-pointer-memaccess): Document enhancement.
(-Wstringop-truncation): Document new option.

gcc/ada/ChangeLog:

PR c/81117
* ada/adadecode.c (__gnat_decode): Use memcpy instead of strncpy.
* ada/argv.c (__gnat_fill_arg, __gnat_fill_env): Same.

gcc/c-family/ChangeLog:

PR c/81117
* c-common.c (catenate_strings): Use memcpy instead of strncpy.
* c-warn.c (sizeof_pointer_memaccess_warning): Handle arrays.
* c.opt (-Wstringop-truncation): New option.

gcc/fortran/ChangeLog:

PR c/81117
* gcc/fortran/decl.c (build_sym): Use strcpy instead of strncpy.

gcc/objc/ChangeLog:

PR c/81117
* objc-encoding.c (encode_type): Use memcpy instead of strncpy.

gcc/testsuite/ChangeLog:

PR c/81117
* c-c++-common/Wsizeof-pointer-memaccess3.c: New test.
* c-c++-common/Wstringop-overflow.c: Same.
* c-c++-common/Wstringop-truncation.c: Same.
* c-c++-common/Wsizeof-pointer-memaccess2.c: Adjust.
* c-c++-common/attr-nonstring-2.c: New test.
* g++.dg/torture/Wsizeof-pointer-memaccess1.C: Adjust.
* g++.dg/torture/Wsizeof-pointer-memaccess2.C: Same.
* gcc.dg/torture/pr63554.c: Same.
* gcc.dg/Walloca-1.c: Disable macro tracking.

From-SVN: r254630

GCOV: create one intermediate file per a gcno file (PR gcov-profile/82702).

2017-11-10  Martin Liska  <mliska@suse.cz>

PR gcov-profile/82702
* gcov.c (main): Handle intermediate files in a different
way.
(get_gcov_intermediate_filename): New function.
(output_gcov_file): Remove support of intermediate files.
(generate_results): Allocate intermediate file.
(release_structures): Clean-up properly fn_end.
(output_intermediate_file): Start iterating with line 1.

From-SVN: r254629

re PR tree-optimization/82929 (r254579 causes ICE: tree check: expected ssa_name, have array_ref in has_single_use, at ssa-iterators.h:400)

PR tree-optimization/82929
* gimple-ssa-store-merging.c (struct store_immediate_info): Add
ops_swapped_p non-static data member.
(store_immediate_info::store_immediate_info): Clear it.
(imm_store_chain_info::coalesce_immediate_stores): If swapping
ops set ops_swapped_p.
(count_multiple_uses): Handle ops_swapped_p.

* gcc.dg/pr82929.c: New test.
* g++.dg/opt/pr82929.C: New test.

From-SVN: r254628

GCOV: do not support unexecuted blocks in Ada

2017-11-10  Martin Liska  <mliska@suse.cz>

* coverage.c (coverage_init): Stream information about
support of has_unexecuted_blocks.
* doc/gcov.texi: Document that.
* gcov-dump.c (dump_gcov_file): Support it in gcov_dump tool.
* gcov.c (read_graph_file): Likewise.
(output_line_beginning): Fix a small issue with
color output.

From-SVN: r254627

[ARM,testsuite] Skip copysign_softfloat_1.c on hard-float targets

gcc/testsuite/

2017-11-10  Christophe Lyon  <christophe.lyon@linaro.org>

* lib/target-supports.exp (check_effective_target_arm_soft_ok):
New function.
* gcc.target/arm/copysign_softfloat_1.c: Require arm_soft_ok
effective target.

From-SVN: r254626

tree-predcom.c (determine_roots_comp): Avoid memory leak by freeing reference of trivial component.

* tree-predcom.c (determine_roots_comp): Avoid memory leak by freeing
reference of trivial component.

From-SVN: r254625

re PR fortran/82934 (Segfault on assumed character length in allocate)

2017-11-10  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/82934
* trans-stmt.c (gfc_trans_allocate): Remove the gcc_assert on
null string length for assumed length typespec and set
expr3_esize to NULL_TREE;

2017-11-10  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/82934
* gfortran.dg/allocate_assumed_charlen_1.f90: New test.

From-SVN: r254624

re PR bootstrap/82916 (gcc miscompiled during stagefeedback (PGO bootstrap))

PR bootstrap/82916
* gimple-ssa-store-merging.c
(pass_store_merging::terminate_all_aliasing_chains): For
gimple_store_p stmts also call refs_output_dependent_p.

* gcc.dg/store_merging_2.c: Only expect 2 successful mergings instead
of 3.
* gcc.dg/pr82916.c: New test.

From-SVN: r254623

[testsuite] Fix retrieval of testname

When gcc-dg-runtest is used to run a test the test is run several times
with different options. For clarity of the log, the test infrastructure
then append the options to the testname. This means that all the code
that must deal with the testcase itself (eg. removing the output files
after the test has run) needs to remove the option name.

There is already a pattern (see below) for this in several place of the
testsuite framework but it is also missing in many places. This patch
fixes all of these places. The pattern is as follows:

set testcase [testname-for-summary]
; The name might include a list of options; extract the file name.
set testcase [lindex $testcase 0]

2017-11-10  Thomas Preud'homme  <thomas.preudhomme@arm.com>

    gcc/testsuite/
    * lib/scanasm.exp (scan-assembler): Extract filename from testname used
    in summary.
    (scan-assembler-not): Likewise.
    (scan-hidden): Likewise.
    (scan-not-hidden): Likewise.
    (scan-stack-usage): Likewise.
    (scan-stack-usage-not): Likewise.
    (scan-assembler-times): Likewise.
    (scan-assembler-dem): Likewise.
    (scan-assembler-dem-not): Likewise.
    (object-size): Likewise.
    (scan-lto-assembler): Likewise.
    * lib/scandump.exp (scan-dump): Likewise.
    (scan-dump-times): Likewise.
    (scan-dump-not): Likewise.
    (scan-dump-dem): Likewise.
    (scan-dump-dem-not): Likewise

From-SVN: r254622

Update copyright years

From-SVN: r254621

utils.c (convert): Add comment and do not fall through to the next case.

* gcc-interface/utils.c (convert) <RECORD_TYPE>: Add comment and do
not fall through to the next case.
<ARRAY_TYPE>: Deal specially with a dereference from another array
type with the same element type.

From-SVN: r254618

Daily bump.

From-SVN: r254617

re PR rtl-optimization/82913 (ICE: Segmentation fault in try_merge_compare)

PR rtl-optimization/82913
* compare-elim.c (try_merge_compare): Punt if def_insn is not
single set.

* gcc.c-torture/compile/pr82913.c: New test.

From-SVN: r254614

vr-values.h: New file with vr_values class.

* vr-values.h: New file with vr_values class.
* tree-vrp.c: Include vr-values.h
(vrp_value_range_pool, vrp_equiv_obstack, num_vr_values): Move static
data objects into the vr_values class.
(vr_value, values_propagated, vr_phi_edge_counts): Likewise.
(get_value_range): Make it a member function within vr_values class.
(set_defs_to_varying, update_value_range, add_equivalence): Likewise.
(vrp_stmt_computes_nonzero_p, op_with_boolean_value_range_p): Likewise.
(op_with_constant_singleton_value_range): Likewise.
(extract_range_for_var_from_comparison_expr): Likewise.
(extract_range_from_assert, extract_range_from_ssa_name): Likewise.
(extract_range_from_binary_expr): Likewise.
(extract_range_from_unary_expr): Likewise.
(extract_range_from_cond_expr, extrat_range_from_comparison): Likewise.
(check_for_binary_op_overflow, extract_range_basic): Likewise.
(extract_range_from_assignment, adjust_range_with_scev): Likewise.
(dump_all_value_ranges, get_vr_for_comparison): Likewise.
(compare_name_with_value, compare_names): Likewise.
(vrp_evaluate_conditional_warnv_with_ops_using_ranges): Likewise.
(vrp_evaluate_conditional_warnv_with_ops): Likewise.  Remove prototype.
(vrp_evaluate_conditional, vrp_visit_cond_stmt): Likewise.
(vrp_visit_switch_stmt, extract_range_from_stmt): Likewise.
(extract_range_from_phi_node): Likewise.
(simplify_truth_ops_using_ranges): Likewise.
(simplify_div_or_mod_using_ranges): Likewise.
(simplify_min_or_max_using_ranges, simplify_abs_using_ranges): Likewise.
(simplify_bit_ops_using_ranges, simplify_cond_using_ranges_1): Likewise.
(simplify_cond_using_ranges_2, simplify_switch_using_ranges): Likewise.
(simplify_float_conversion_using_ranges): Likewise.
(simplify_internal_call_using_ranges): Likewise.
(two_valued_val_range_p, simplify_stmt_using_ranges): Likewise.
(vrp_visit_assignment_or_call): Likewise.  Smuggle class instance
poitner via x_vr_values for calls into gimple folder.
(vrp_initialize_lattice): Make this the vr_values ctor.
(vrp_free_lattice): Make this the vr_values dtor.
(set_vr_value): New function.
(class vrp_prop): Add vr_values data member.  Add various member
functions as well as member functions that delegate to vr_values.
(check_array_ref): Make a member function within vrp_prop class.
(search_for_addr_array, vrp_initialize): Likewise.
(vrp_finalize): Likewise.  Revamp to avoid direct access to
vr_value, values_propagated, etc.
(check_array_bounds): Extract vrp_prop class instance pointer from
walk info structure.  Use it to call member functions.
(check_all_array_refs): Make a member function within vrp_prop class.
Smuggle class instance pointer via walk info structure.
(x_vr_values): New local static.
(vrp_valueize): Use x_vr_values to get class instance.
(vr_valueize_1): Likewise.
(class vrp_folder): Add vr_values data member.  Add various member
functions as well as member functions that delegate to vr_values.
(fold_predicate_in): Make a mber fucntion within vrp_folder class.
(simplify_stmt_for_jump_threading): Extract smuggled vr_values
class instance from vr_values.  Use it to call member functions.
(vrp_dom_walker): Add vr_values data member.
(vrp_dom_walker::after_dom_children): Smuggle vr_values class
instance via x_vr_values.
(identify_jump_threads): Accept vr_values as argument.  Store
it into the walker structure.
(evrp_dom_walker): Add vr_values class data member.  Add various
delegators.
(evrp_dom_walker::try_find_new_range): Use vr_values data
member to access the memory allocator.
(evrp_dom_walker::before_dom_children): Store vr_values class
instance into the vrp_folder class.
(evrp_dom_walker::push_value_range): Rework to avoid direct
access to num_vr_values and vr_value.
(evrp_dom_walker::pop_value_range): Likewise.
(execute_early_vrp): Remove call to vrp_initialize_lattice.
Use vr_values to get to dump_all_value_ranges member function.
Remove call to vrp_free_lattice.  Call vrp_initialize, vrp_finalize,
and simplify_cond_using_ranges_2 via vrp_prop class instance.
Pass vr_values class instance down to identify_jump_threads.
Remove call to vrp_free_lattice.
(debug_all_value_ranges): Remove.

From-SVN: r254613

tree-vrp.c (vrp_prop): Move class to earlier point in the file.

* tree-vrp.c (vrp_prop): Move class to earlier point in the file.
(vrp_folder): Likewise.

From-SVN: r254612

tree-vrp.c (set_value_range): Do not reference vrp_equiv_obstack.

* tree-vrp.c (set_value_range): Do not reference vrp_equiv_obstack.
Get it from the existing bitmap instead.
(vrp_intersect_ranges_1): Likewise.

From-SVN: r254611

sync/atomic, runtime/internal/atomic: don't assume reads from 0 fail

    For a misaligned address force a panic rather than assuming that reading
    from the address 0 will cause one.

    Reviewed-on: https://go-review.googlesource.com/69850

From-SVN: r254610

gimple-ssa-store-merging.c (struct store_immediate_info): Add bit_not_p field.

* gimple-ssa-store-merging.c (struct store_immediate_info): Add
bit_not_p field.
(store_immediate_info::store_immediate_info): Add bitnotp argument,
set bit_not_p to it.
(imm_store_chain_info::coalesce_immediate_stores): Break group
if bit_not_p is different.
(count_multiple_uses, split_group,
imm_store_chain_info::output_merged_store): Handle info->bit_not_p.
(handled_load): Avoid multiple chained BIT_NOT_EXPRs.
(pass_store_merging::process_store): Handle BIT_{AND,IOR,XOR}_EXPR
result inverted using BIT_NOT_EXPR, compute bit_not_p, pass it
to store_immediate_info ctor.

From-SVN: r254606

re PR fortran/78619 (ICE in copy_reference_ops_from_ref, at tree-ssa-sccvn.c:889)

2017-11-09  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/78619
* check.c (same_type_check): Introduce a new argument 'assoc'
with default value false. If this is true, use the symbol type
spec of BT_PROCEDURE expressions.
(gfc_check_associated): Set 'assoc' true in the call to
'same_type_check'.

2017-11-09  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/78619
* gfortran.dg/pr78619.f90: New test.

From-SVN: r254605

re PR fortran/78814 (ICE in symbol_rank, at fortran/interface.c:1265)

2017-11-09  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/78814
* interface.c (symbol_rank): Check for NULL pointer.

2017-11-09  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/78814
* gfortran.dg/interface_40.f90: New testcase.

From-SVN: r254604

Remove obsolete ECOFF support.

* collect2.c (OBJECT_FORMAT_COFF): Remove EXTENDED_COFF support.
(scan_prog_file): Likewise.

From-SVN: r254603

bb-reorder.c (max_entry_frequency): Remove.

* bb-reorder.c (max_entry_frequency): Remove.
(find_traces, rotate_loop, mark_bb_visited, connect_better_edge_p,
connect_traces, push_to_next_round_p): Remove prototypes.
(find_traces_1_round): Use counts only.
(push_to_next_round_p): Likewise.
(find_traces): Likewise.
(rotate_loop): Likewise.
(find_traces_1_round): Likewise.
(connect_traces): Likewise.
(edge_order): Likewise.

From-SVN: r254602

[ARM] Fix cmse_nonsecure_entry return insn size

A number of instructions are output in assembler form by
output_return_instruction () when compiling a function with the
cmse_nonsecure_entry attribute for Armv8-M Mainline with hardfloat float
ABI. However, the corresponding thumb2_cmse_entry_return insn pattern
does not account for all these instructions in its computing of the
length of the instruction.

This may lead GCC to use the wrong branching instruction due to
incorrect computation of the offset between the branch instruction's
address and the target address.

This commit fixes the mismatch between what output_return_instruction ()
does and what the pattern think it does and adds a note warning about
mismatch in the affected functions' heading comments to ensure code does
not get out of sync again.

Note: no test is provided because the C testcase is fragile (only works
on GCC 6) and the extracted RTL test fails to compile due to bugs in the
RTL frontend (PR82815 and PR82817)

2017-11-09  Thomas Preud'homme  <thomas.preudhomme@arm.com>

    gcc/
    * config/arm/arm.c (output_return_instruction): Add comments to
    indicate requirement for cmse_nonsecure_entry return to account
    for the size of clearing instruction output here.
    (thumb_exit): Likewise.
    * config/arm/thumb2.md (thumb2_cmse_entry_return): Fix length for
    return in hardfloat mode.

From-SVN: r254601

controlled2.adb, [...]: Disable all warnings.

* gnat.dg/controlled2.adb, gnat.dg/controlled4.adb,
gnat.dg/finalized.adb: Disable all warnings.

From-SVN: r254600

rs6000: Separate shrink-wrapping for the TOC register

This makes the TOC register save a component.  If -msave-toc-indirect
is not explicitly disabled, it enables it, and then moves the prologue
code generated for that to a better place.  So far this only matters
for indirect calls (for direct calls the save is done in the PLT stub).
The restore is always done directly after the bl insn (the compiler
generates a nop there, the linker replaces it with a load).

* config/rs6000/rs6000.c (machine_function): Add a bool,
"toc_is_wrapped_separately".
(rs6000_option_override_internal): Enable OPTION_MASK_SAVE_TOC_INDIRECT
if it wasn't explicitly set or unset, we are optimizing for speed, and
doing separate shrink-wrapping.
(rs6000_get_separate_components): Enable the TOC component if
saving the TOC register in the prologue.
(rs6000_components_for_bb): Handle the TOC component.
(rs6000_emit_prologue_components): Store the TOC register where needed.
(rs6000_set_handled_components): Mark TOC as handled, if handled.
(rs6000_emit_prologue): Don't save the TOC if that is already done.

From-SVN: r254599

Moving parameter manipulation into its own file

2017-11-09  Martin Jambor  <mjambor@suse.cz>

* ipa-param-manipulation.c: New file.
* ipa-param-manipulation.h: Likewise.
* Makefile.in (OBJS): Add ipa-param-manipulation.o.
(PLUGIN_HEADERS): Addded ipa-param-manipulation.h
* ipa-param.h (ipa_parm_op): Moved to ipa-param-manipulation.h.
(ipa_parm_adjustment): Likewise.
(ipa_parm_adjustment_vec): Likewise.
(ipa_get_vector_of_formal_parms): Moved declaration to
ipa-param-manipulation.h.
(ipa_get_vector_of_formal_parm_types): Likewise.
(ipa_modify_formal_parameters): Likewise.
(ipa_modify_call_arguments): Likewise.
(ipa_combine_adjustments): Likewise.
(ipa_dump_param_adjustments): Likewise.
(ipa_modify_expr): Likewise.
(ipa_get_adjustment_candidate): Likewise.
* ipa-prop.c (ipa_get_vector_of_formal_parms): Moved to
ipa-param-manipulation.c.
(ipa_get_vector_of_formal_parm_types): Likewise.
(ipa_modify_formal_parameters): Likewise.
(ipa_modify_call_arguments): Likewise.
(ipa_modify_expr): Likewise.
(get_ssa_base_param): Likewise.
(ipa_get_adjustment_candidate): Likewise.
(index_in_adjustments_multiple_times_p): Likewise.
(ipa_combine_adjustments): Likewise.
(ipa_dump_param_adjustments): Likewise.
* tree-sra.c: Also include ipa-param-manipulation.h
* omp-simd-clone.c: Include ipa-param-manipulation.h instead of
ipa-param.h.

From-SVN: r254598

Add a vect_masked_store target selector

This patch adds a target selector that says whether the target
supports IFN_MASK_STORE.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/sourcebuild.texi (vect_masked_store): Document.

gcc/testsuite/
* lib/target-supports.exp (check_effective_target_vect_masked_store):
New proc.
* gcc.dg/vect/vect-cselim-1.c (foo): Mention that the second loop
is vectorizable with masked stores.  Update scan-tree-dump-times
accordingly.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254597

Add a vect_align_stack_vars target selector

This patch adds a target selector to say whether it's possible to
align a local variable to the target's preferred vector alignment.
This can be false for large vectors if the alignment is only
a preference and not a hard requirement (and thus if there is no
need to support a stack realignment mechanism).

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/sourcebuild.texi (vect_align_stack_vars): Document.

gcc/testsuite/
* lib/target-supports.exp
(check_effective_target_vect_align_stack_vars): New proc.
* gcc.dg/vect/vect-23.c: Only expect the array to be aligned if
vect_align_stack_vars.
* gcc.dg/vect/vect-24.c: Likewise.
* gcc.dg/vect/vect-25.c: Likewise.
* gcc.dg/vect/vect-26.c: Likewise.
* gcc.dg/vect/vect-32-big-array.c: Likewise.
* gcc.dg/vect/vect-32.c: Likewise.
* gcc.dg/vect/vect-40.c: Likewise.
* gcc.dg/vect/vect-42.c: Likewise.
* gcc.dg/vect/vect-46.c: Likewise.
* gcc.dg/vect/vect-48.c: Likewise.
* gcc.dg/vect/vect-52.c: Likewise.
* gcc.dg/vect/vect-54.c: Likewise.
* gcc.dg/vect/vect-62.c: Likewise.
* gcc.dg/vect/vect-67.c: Likewise.
* gcc.dg/vect/vect-75-big-array.c: Likewise.
* gcc.dg/vect/vect-75.c: Likewise.
* gcc.dg/vect/vect-77-alignchecks.c: Likewise.
* gcc.dg/vect/vect-78-alignchecks.c: Likewise.
* gcc.dg/vect/vect-89-big-array.c: Likewise.
* gcc.dg/vect/vect-89.c: Likewise.
* gcc.dg/vect/vect-96.c: Likewise.
* gcc.dg/vect/vect-multitypes-3.c: Likewise.
* gcc.dg/vect/vect-multitypes-6.c: Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254596

Add a vect_variable_length target selector

This patch adds a target selector for variable-length vectors.
Initially it's always false, but the SVE patch provides a case
in which it's true.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/sourcebuild.texi (vect_variable_length): Document.

gcc/testsuite/
* lib/target-supports.exp
(check_effective_target_vect_variable_length): New proc.
* gcc.dg/vect/pr60482.c: XFAIL test for no epilog loop if
vect_variable_length.
* gcc.dg/vect/slp-reduc-6.c: XFAIL two-operation SLP if
vect_variable_length.
* gcc.dg/vect/vect-alias-check-5.c: XFAIL alias optimization if
vect_variable_length.
* gfortran.dg/vect/fast-math-mgrid-resid.f: XFAIL predictive
commoning optimization if vect_variable_length.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254595

Add a vect_unaligned_possible target selector

This patch adds a target selector that says whether we can ever
generate an "unaligned" accesses, where "unaligned" is relative
to the target's preferred vector alignment.  This is already true if:

   vect_no_align && { ! vect_hw_misalign }

i.e. if the target doesn't have any alignment mechanism and also
doesn't allow unaligned accesses.  It is also true (for the things
tested by gcc.dg/vect) if the target only wants things to be aligned
to an element; in that case every normal scalar access is "vector aligned".

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/sourcebuild.texi (vect_unaligned_possible): Document.

gcc/testsuite/
* lib/target-supports.exp
(check_effective_target_vect_unaligned_possible): New proc.
* gcc.dg/vect/slp-25.c: Extend XFAIL of peeling for alignment from
vect_no_align && { ! vect_hw_misalign } to ! vect_unaligned_possible.
* gcc.dg/vect/vect-multitypes-1.c: Likewise.
* gcc.dg/vect/vect-109.c: XFAIL vectorisation of an unaligned
access to ! vect_unaligned_possible.
* gcc.dg/vect/vect-33.c: Likewise.
* gcc.dg/vect/vect-42.c: Likewise.
* gcc.dg/vect/vect-56.c: Likewise.
* gcc.dg/vect/vect-60.c: Likewise.
* gcc.dg/vect/vect-96.c: Likewise.
* gcc.dg/vect/vect-peel-1.c: Likewise.
* gcc.dg/vect/vect-27.c: Extend XFAIL of unaligned vectorization from
vect_no_align && { ! vect_hw_misalign } to ! vect_unaligned_possible.
* gcc.dg/vect/vect-29.c: Likewise.
* gcc.dg/vect/vect-44.c: Likewise.
* gcc.dg/vect/vect-48.c: Likewise.
* gcc.dg/vect/vect-50.c: Likewise.
* gcc.dg/vect/vect-52.c: Likewise.
* gcc.dg/vect/vect-72.c: Likewise.
* gcc.dg/vect/vect-75-big-array.c: Likewise.
* gcc.dg/vect/vect-75.c: Likewise.
* gcc.dg/vect/vect-77-alignchecks.c: Likewise.
* gcc.dg/vect/vect-77-global.c: Likewise.
* gcc.dg/vect/vect-78-alignchecks.c: Likewise.
* gcc.dg/vect/vect-78-global.c: Likewise.
* gcc.dg/vect/vect-multitypes-3.c: Likewise.
* gcc.dg/vect/vect-multitypes-4.c: Likewise.
* gcc.dg/vect/vect-multitypes-6.c: Likewise.
* gcc.dg/vect/vect-peel-4.c: Likewise.
* gcc.dg/vect/vect-peel-3.c: Likewise, and also for peeling
for alignment.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254594

Add a vect_element_align_preferred target selector

This patch adds a target selector for targets whose
preferred_vector_alignment is the alignment of one element.  We'll never
peel in that case, and the step of a loop that operates on normal (as
opposed to packed) elements will always divide the preferred alignment.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/sourcebuild.texi (vect_element_align_preferred): Document.

gcc/testsuite/
* lib/target-supports.exp
(check_effective_target_vect_element_align_preferred): New proc.
(check_effective_target_vect_peeling_profitable): Test it.
* gcc.dg/vect/no-section-anchors-vect-31.c: Don't expect peeling
if vect_element_align_preferred.
* gcc.dg/vect/no-section-anchors-vect-64.c: Likewise.
* gcc.dg/vect/pr65310.c: Likewise.
* gcc.dg/vect/vect-26.c: Likewise.
* gcc.dg/vect/vect-54.c: Likewise.
* gcc.dg/vect/vect-56.c: Likewise.
* gcc.dg/vect/vect-58.c: Likewise.
* gcc.dg/vect/vect-60.c: Likewise.
* gcc.dg/vect/vect-89-big-array.c: Likewise.
* gcc.dg/vect/vect-89.c: Likewise.
* gcc.dg/vect/vect-92.c: Likewise.
* gcc.dg/vect/vect-peel-1.c: Likewise.
* gcc.dg/vect/vect-outer-3a-big-array.c: Expect the step to
divide the alignment if vect_element_align_preferred.
* gcc.dg/vect/vect-outer-3a.c: Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254593

Add vect_perm3_* target selectors

SLP load permutation fails if any individual permutation requires more
than two vector inputs.  For 128-bit vectors, it's possible to permute
3 contiguous loads of 32-bit and 8-bit elements, but not 16-bit elements
or 64-bit elements.  The results are reversed for 256-bit vectors,
and so on for wider vectors.

This patch adds a routine that tests whether a permute will require
three vectors for a given vector count and element size, then adds
vect_perm3_* target selectors for the cases that we currently use.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/sourcebuild.texi (vect_perm_short, vect_perm_byte): Document
previously undocumented selectors.
(vect_perm3_byte, vect_perm3_short, vect_perm3_int): Document.

gcc/testsuite/
* lib/target-supports.exp (vect_perm_supported): New proc.
(check_effective_target_vect_perm3_int): Likewise.
(check_effective_target_vect_perm3_short): Likewise.
(check_effective_target_vect_perm3_byte): Likewise.
* gcc.dg/vect/slp-perm-1.c: Expect SLP load permutation to
succeed if vect_perm3_int.
* gcc.dg/vect/slp-perm-5.c: Likewise.
* gcc.dg/vect/slp-perm-6.c: Likewise.
* gcc.dg/vect/slp-perm-7.c: Likewise.
* gcc.dg/vect/slp-perm-8.c: Likewise vect_perm3_byte.
* gcc.dg/vect/slp-perm-9.c: Likewise vect_perm3_short.
Use vect_perm_short instead of vect_perm.  Add a scan-tree-dump-not
test for vect_perm3_short targets.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254592

Don't assume vect_multiple_sizes means 2 sizes

Some tests assumed that there would only be 2 vector sizes if
vect_multiple_sizes, whereas for SVE there are three (SVE, 128-bit
and 64-bit).  This patch replaces scan-tree-dump-times with
scan-tree-dump for vect_multiple_sizes but keeps it for
!vect_multiple_sizes.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/testsuite/
* gcc.dg/vect/no-vfa-vect-101.c: Use scan-tree-dump rather than
scan-tree-dump-times for vect_multiple_sizes.
* gcc.dg/vect/no-vfa-vect-102.c: Likewise.
* gcc.dg/vect/no-vfa-vect-102a.c: Likewise.
* gcc.dg/vect/no-vfa-vect-37.c: Likewise.
* gcc.dg/vect/no-vfa-vect-79.c: Likewise.
* gcc.dg/vect/vect-104.c: Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254591

Add available_vector_sizes to target-supports.exp

This patch adds a routine that lists the available vector sizes
for a target and uses it for some existing target conditions.
Later patches add more uses.

The cases are taken from multiple_sizes.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/testsuite/
* lib/target-supports.exp (available_vector_sizes): New proc.
(check_effective_target_vect_multiple_sizes): Use it.
(check_effective_target_vect64): Likewise.
(check_effective_target_vect_sizes_32B_16B): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254590

Add VECTOR_BITS to tree-vect.h

Several vector tests are sensitive to the vector size.  This patch adds
a VECTOR_BITS macro to tree-vect.h to select the expected vector size
and uses it to influence iteration counts and array sizes.  The tests
keep the original values if the vector size is small enough.

For now VECTOR_BITS is always 128, but the SVE patches add other values.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/testsuite/
* gcc.dg/vect/tree-vect.h (VECTOR_BITS): Define.
* gcc.dg/vect/bb-slp-pr69907.c: Include tree-vect.h.
(N): New macro.
(foo): Use it instead of hard-coded 320.
* gcc.dg/vect/no-scevccp-outer-7.c (N): Redefine if the default
value is too small for VECTOR_BITS.
* gcc.dg/vect/no-scevccp-vect-iv-3.c (N): Likewise.
* gcc.dg/vect/no-section-anchors-vect-31.c (N): Likewise.
* gcc.dg/vect/no-section-anchors-vect-36.c (N): Likewise.
* gcc.dg/vect/slp-perm-9.c (N): Likewise.
* gcc.dg/vect/vect-32.c (N): Likewise.
* gcc.dg/vect/vect-75.c (N, OFF): Likewise.
* gcc.dg/vect/vect-77-alignchecks.c (N, OFF): Likewise.
* gcc.dg/vect/vect-78-alignchecks.c (N, OFF): Likewise.
* gcc.dg/vect/vect-89.c (N): Likewise.
* gcc.dg/vect/vect-96.c (N): Likewise.
* gcc.dg/vect/vect-multitypes-3.c (N): Likewise.
* gcc.dg/vect/vect-multitypes-6.c (N): Likewise.
* gcc.dg/vect/vect-over-widen-1.c (N): Likewise.
* gcc.dg/vect/vect-over-widen-4.c (N): Likewise.
* gcc.dg/vect/vect-reduc-pattern-1a.c (N): Likewise.
* gcc.dg/vect/vect-reduc-pattern-1b.c (N): Likewise.
* gcc.dg/vect/vect-reduc-pattern-2a.c (N): Likewise.
* gcc.dg/vect/no-section-anchors-vect-64.c (NINTS): New macro.
(N): Redefine in terms of NINTS.
(ia, ib, ic): Use NINTS instead of hard-coded constants in the
array bounds.
* gcc.dg/vect/no-section-anchors-vect-69.c (NINTS): New macro.
(N): Redefine in terms of NINTS.
(test1): Replace a and b fields with NINTS - 2 ints of padding.
(main1): Use NINTS instead of hard-coded constants.
* gcc.dg/vect/section-anchors-vect-69.c (NINTS): New macro.
(N): Redefine in terms of NINTS.
(test1): Replace a and b fields with NINTS - 2 ints of padding.
(test2): Remove incorrect comments about alignment.
(main1): Use NINTS instead of hard-coded constants.
* gcc.dg/vect/pr45752.c (N): Redefine if the default value is
too small for VECTOR_BITS.
(main): Continue to use canned results for the default value of N,
but compute the expected results from scratch for other values.
* gcc.dg/vect/slp-perm-1.c (N, main): As for pr45752.c.
* gcc.dg/vect/slp-perm-4.c (N, main): Likewise.
* gcc.dg/vect/slp-perm-5.c (N, main): Likewise.
* gcc.dg/vect/slp-perm-6.c (N, main): Likewise.
* gcc.dg/vect/slp-perm-7.c (N, main): Likewise.
* gcc.dg/vect/pr65518.c (NINTS, N, RESULT): New macros.
(giga): Use NINTS as the array bound.
(main): Use NINTS, N and RESULT.
* gcc.dg/vect/pr65947-5.c (N): Redefine if the default value is
too small for VECTOR_BITS.
(main): Fill in any remaining elements of A programmatically.
* gcc.dg/vect/pr81136.c: Include tree-vect.h.
(a): Use VECTOR_BITS to set the alignment of the target structure.
* gcc.dg/vect/slp-19c.c (N): Redefine if the default value is
too small for VECTOR_BITS.
(main1): Continue to use the canned input for the default value of N,
but compute the input from scratch for other values.
* gcc.dg/vect/slp-28.c (N): Redefine if the default value is
too small for VECTOR_BITS.
(in1, in2, in3): Remove initialization.
(check1, check2): Delete.
(main1): Initialize in1, in2 and in3 here.  Check every element
of the vectors and compute the expected values directly instead
of using an array.
* gcc.dg/vect/slp-perm-8.c (N): Redefine if the default value is
too small for VECTOR_BITS.
(foo, main): Change type of "i" to int.
* gcc.dg/vect/vect-103.c (NINTS): New macro.
(N): Redefine in terms of N.
(c): Delete.
(main1): Use NINTS.  Check the result from a and b directly.
* gcc.dg/vect/vect-67.c (NINTS): New macro.
(N): Redefine in terms of N.
(main1): Use NINTS for the inner array bounds.
* gcc.dg/vect/vect-70.c (NINTS, OUTERN): New macros.
(N): Redefine in terms of NINTS.
(s): Keep the outer dimensions as 4 even if N is larger than 24.
(tmp1): New variable.
(main1): Only define a local tmp1 if NINTS is relatively small.
Use OUTERN for the outer loops and NINTS for the inner loops.
* gcc.dg/vect/vect-91.c (OFF): New macro.
(a, main3): Use it.
* gcc.dg/vect/vect-92.c (NITER): New macro.
(main1, main2): Use it.
* gcc.dg/vect/vect-93.c (N): Rename to...
(N1): ...this.
(main): Update accordingly.
(N2): New macro.
(main1): Use N1 instead of 3001 and N2 insteaed of 10.
* gcc.dg/vect/vect-multitypes-1.c (NSHORTS, NINTS): New macros.
(N): Redefine in terms of NSHORTS.
(main1): Use NINTS - 1 instead of 3 and NSHORTS - 1 instead of 7.
(main): Likewise.
* gcc.dg/vect/vect-over-widen-3-big-array.c (N): Define to VECTOR_BITS.
(foo): Truncate the expected value to the type of *d.
* gcc.dg/vect/vect-peel-3.c (NINTS, EXTRA): New macros.
(ia, ib, ic, main): Use EXTRA.
(main): Use NINTS.
(RES_A, RES_B, REC_C): New macros.
(RES): Redefine as their sum.
* gcc.dg/vect/vect-reduc-or_1.c (N): New macro.
(in): Change number of elements to N.
(main): Update accordingly.  Calculate the expected result.
* gcc.dg/vect/vect-reduc-or_2.c (N, in, main): As for
vect-reduc-or-1.c.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254589

Consistently use asm volatile ("" ::: "memory") in vect tests

The vectoriser tests used a combination of:

1) if (impossible condition) abort ();
2) volatile int x; ... *x = ...;
3) asm volatile ("" ::: "memory");

to prevent vectorisation of a set-up loop.  The problem with 1) is that
the compiler can often tell that the condition is false and optimise
it away before vectorisation.

This was already happening in slp-perm-9.c, which is why the test was
expecting one loop to be vectorised even when the required permutes
weren't supported.  It becomes a bigger problem with SVE, which is
able to vectorise more set-up loops.

The point of this patch is therefore to replace 1) with something else.
2) should work most of the time, but we don't usually treat non-volatile
accesses as aliasing unrelated volatile accesses, so I think in principle
we could split the loop into one that does the set-up and one that does
the volatile accesses.  3) seems more robust because it's also a wild
read and write.

The patch therefore tries to replace all instances of 1) and 2) with 3).

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/testsuite/
* gcc.dg/vect/bb-slp-cond-1.c (main): Add an asm volatile
to the set-up loop.
* gcc.dg/vect/slp-perm-7.c (main): Prevent vectorisation with
asm volatile ("" ::: "memory") instead of a conditional abort.
Update the expected vector loop count accordingly.
* gcc.dg/vect/slp-perm-9.c (main): Likewise.
* gcc.dg/vect/bb-slp-1.c (main1): Prevent vectorisation with
asm volatile ("" ::: "memory") instead of a conditional abort.
* gcc.dg/vect/slp-23.c (main): Likewise,
* gcc.dg/vect/slp-35.c (main): Likewise,
* gcc.dg/vect/slp-37.c (main): Likewise,
* gcc.dg/vect/slp-perm-4.c (main): Likewise.
* gcc.dg/vect/bb-slp-24.c (foo): Likewise.  Remove dummy argument.
(main): Update call accordingly.
* gcc.dg/vect/bb-slp-25.c (foo, main): As for bb-slp-24.c.
* gcc.dg/vect/bb-slp-26.c (foo, main): Likewise.
* gcc.dg/vect/bb-slp-29.c (foo, main): Likewise.
* gcc.dg/vect/no-vfa-vect-102.c (foo): Delete.
(main): Don't initialize it.
(main1): Prevent vectorisation with asm volatile ("" ::: "memory")
instead of a conditional abort.
* gcc.dg/vect/no-vfa-vect-102a.c (foo, main1, main): As for
no-vfa-vect-102.c
* gcc.dg/vect/vect-103.c (foo, main1, main): Likewise.
* gcc.dg/vect/vect-104.c (foo, main1, main): Likewise.
* gcc.dg/vect/pr42709.c (main1): Remove dummy argument.
Prevent vectorisation with asm volatile ("" ::: "memory")
instead of a conditional abort.
* gcc.dg/vect/slp-13-big-array.c (y): Delete.
(main1): Prevent vectorisation with asm volatile ("" ::: "memory")
instead of a conditional abort.
* gcc.dg/vect/slp-3-big-array.c (y, main1): As for slp-13-big-array.c.
* gcc.dg/vect/slp-34-big-array.c (y, main1): Likewise.
* gcc.dg/vect/slp-4-big-array.c (y, main1): Likewise.
* gcc.dg/vect/slp-multitypes-11-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-105.c (y, main1): Likewise.
* gcc.dg/vect/vect-105-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-112-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-15-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-2-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-34-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-6-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-73-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-74-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-75-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-76-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-80-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-97-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-all-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-reduc-1char-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-reduc-2char-big-array.c (y, main1): Likewise.
* gcc.dg/vect/vect-strided-a-mult.c (y, main1): Likewise.
* gcc.dg/vect/vect-strided-a-u16-i2.c (y, main1): Likewise.
* gcc.dg/vect/vect-strided-a-u16-i4.c (y, main1): Likewise.
* gcc.dg/vect/vect-strided-a-u16-mult.c (y, main1): Likewise.
* gcc.dg/vect/vect-strided-a-u8-i2-gap.c (y, main1): Likewise.
* gcc.dg/vect/vect-strided-a-u8-i8-gap2-big-array.c (y, main1):
Likewise.
* gcc.dg/vect/vect-strided-a-u8-i8-gap2.c (y, main1): Likewise.
* gcc.dg/vect/vect-strided-a-u8-i8-gap7-big-array.c (y, main1):
Likewise.
* gcc.dg/vect/vect-strided-a-u8-i8-gap7.c (y, main1): Likewise.
* gcc.dg/vect/slp-24.c (y): Delete.
(main): Prevent vectorisation with asm volatile ("" ::: "memory")
instead of a conditional abort.
* gcc.dg/vect/slp-24-big-array.c (y, main): As for slp-24.c.
* gcc.dg/vect/vect-98-big-array.c (y, main): Likewise.
* gcc.dg/vect/vect-bswap16.c (y, main): Likewise.
* gcc.dg/vect/vect-bswap32.c (y, main): Likewise.
* gcc.dg/vect/vect-bswap64.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-mult-char-ls.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-mult.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-same-dr.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u16-i2.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u16-i4.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u32-i4.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u32-i8.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i2-gap.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i2.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8-gap2-big-array.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8-gap2.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8-gap4-big-array.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8-gap4-unknown.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8-gap4.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8-gap7-big-array.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8-gap7.c (y, main): Likewise.
* gcc.dg/vect/vect-strided-u8-i8.c (y, main): Likewise.
* gcc.dg/vect/vect-10-big-array.c (y): Delete.
(foo): Prevent vectorisation with asm volatile ("" ::: "memory")
instead of a conditional abort.
* gcc.dg/vect/vect-double-reduc-6-big-array.c (y, foo): As for
vect-10-big-array.c.
* gcc.dg/vect/vect-reduc-pattern-1b-big-array.c (y, foo): Likewise.
* gcc.dg/vect/vect-reduc-pattern-1c-big-array.c (y, foo): Likewise.
* gcc.dg/vect/vect-reduc-pattern-2b-big-array.c (y, foo): Likewise.
* gcc.dg/vect/vect-117.c (foo): Delete.
(main): Don't initalize it.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254588

exp_util.adb, freeze.adb: Minor reformatting.

2017-11-09  Gary Dismukes  <dismukes@adacore.com>

* exp_util.adb, freeze.adb: Minor reformatting.

From-SVN: r254587

Be stricter about CONST_VECTOR operands

The recent gen_vec_duplicate patches used CONST_VECTOR for all
constants, but the documentation says:

  @findex const_vector
  @item (const_vector:@var{m} [@var{x0} @var{x1} @dots{}])
  Represents a vector constant.  The square brackets stand for the vector
  containing the constant elements.  @var{x0}, @var{x1} and so on are
  the @code{const_int}, @code{const_double} or @code{const_fixed} elements.

Both the AArch32 and AArch64 ports relied on the elements having
this form and would ICE if the element was something like a CONST
instead.  This showed up as a failure in vect-126.c for both arm-eabi
and aarch64-elf (but not aarch64-linux-gnu, which is what the series
was tested on).

The two obvious options were to redefine CONST_VECTOR to accept all
constants or make gen_vec_duplicate honour the existing documentation.
It looks like other code also assumes that integer CONST_VECTORs contain
CONST_INTs, so the patch does the latter.

I deliberately didn't add an assert to gen_const_vec_duplicate
because it looks like the SPU port *does* expect to be able to create
CONST_VECTORs of symbolic constants.

Also, I think the list above should include const_wide_int for vectors
of TImode and wider.

The new routine takes a mode for consistency with the generators,
and because I think it does make sense to accept all constants for
variable-length:

    (const (vec_duplicate ...))

rather than have some rtxes for which we instead use:

    (vec_duplicate (const ...))

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* doc/rtl.texi (const_vector): Say that elements can be
const_wide_ints too.
* emit-rtl.h (valid_for_const_vec_duplicate_p): Declare.
* emit-rtl.c (valid_for_const_vec_duplicate_p): New function.
(gen_vec_duplicate): Use it instead of CONSTANT_P.
* optabs.c (expand_vector_broadcast): Likewise.

From-SVN: r254586

Improve ivopts handling of forced scales

This patch improves the ivopts address cost calculation for modes
in which an index must be scaled rather than unscaled.  Previously
we would only try the scaled form if the unscaled form was valid.

Many of the SVE tests rely on this when matching scaled indices.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-ssa-loop-ivopts.c (get_address_cost): Try using a
scaled index even if the unscaled address was invalid.
Don't increase the complexity of using a scale in that case.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254585

Base subreg rules on REGMODE_NATURAL_SIZE rather than UNITS_PER_WORD

Originally subregs operated at the word level and subreg offsets
were measured in words.  The offset units were later changed from
words to bytes (SUBREG_WORD became SUBREG_BYTE), but the fundamental
assumption that subregs should operate at the word level remained.
Whether (subreg:M1 (reg:M2 R2) N) is well-formed depended on the
way that M1 and M2 partitioned into words and whether the subword
part of N represented a lowpart.  However, some questions depended
instead on the macro REGMODE_NATURAL_SIZE, which was introduced
as part of the patch that moved from SUBREG_WORD to SUBREG_BYTE.
It is used to decide whether setting (subreg:M1 (reg:M2 R2) N)
clobbers all of R2 or just part of it (df_read_modify_subreg).

Using words doesn't really make sense for modern vector
architectures.  Vector registers are usually bigger than
a word and:

(a) setting the scalar lowpart of them usually clobbers the
    rest of the register (contrary to the subreg rules,
    where only the containing words should be clobbered).

(b) high words of vector registers are often not independently
    addressable, even though that's what the subreg rules expect.

This patch therefore uses REGMODE_NATURAL_SIZE instead of
UNITS_PER_WORD to determine the size of the independently
addressable blocks in an inner register.

This is needed for SVE because the number of words in a vector
mode isn't known at compile time, so isn't a sensible basis
for calculating the number of registers.

The only existing port to define REGMODE_NATURAL_SIZE is
64-bit SPARC, where FP registers are 32 bits.  (This is the
opposite of the use case for SVE, since the natural division
is smaller than a word.)  I compiled the testsuite before and
after the patch for sparc64-linux-gnu and the only test whose
assembly changed was g++.dg/debug/pr65678.C, where the order
of two independent stores was reversed and where a different
register was picked for one pseudo.  The new code was
otherwise equivalent to the old code.

2017-11-09  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/rtl.texi: Rewrite the subreg rules so that they partition
the inner register into REGMODE_NATURAL_SIZE bytes rather than
UNITS_PER_WORD bytes.
* emit-rtl.c (validate_subreg): Divide subregs into blocks
based on REGMODE_NATURAL_SIZE of the inner mode.
(gen_lowpart_common): Split the SCALAR_FLOAT_MODE_P and
!SCALAR_FLOAT_MODE_P cases.  Use REGMODE_NATURAL_SIZE for the latter.
* expmed.c (lowpart_bit_field_p): Divide the value up into
chunks of REGMODE_NATURAL_SIZE rather than UNITS_PER_WORD.
* expr.c (store_constructor): Use REGMODE_NATURAL_SIZE to test
whether something is likely to occupy more than one register.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254583

re PR ipa/82879 (ICE in max, at profile-count.h:889)

PR ipa/82879
* ipa-inline-transform.c (update_noncloned_frequencies): Use
profile_count::adjust_for_ipa_scaling.
* tree-inline.c (copy_bb, copy_cfg_body): Likewise.
* profile-count.c (profile_count::adjust_for_ipa_scaling): New member
function.
* profile-count.h (profile_count::adjust_for_ipa_scaling): Declare.

From-SVN: r254582

[multiple changes]

2017-11-09  Jerome Lambourg  <lambourg@adacore.com>

* gcc-interface/Makefile.in: Add rules to build aarch64-qnx runtimes.

2017-11-09  Hristian Kirtchev  <kirtchev@adacore.com>

* gcc-interface/trans.c (gnat_to_gnu): Add processing for
N_Variable_Reference_Marker nodes.

From-SVN: r254581

[multiple changes]

2017-11-09  Ed Schonberg  <schonberg@adacore.com>

* sem_ch12.adb (Analyze_Generic_Package_Declaration): Handle properly
the pragma Compile_Time_Error when it appears in a generic package
declaration and uses an expanded name to denote the current unit.

2017-11-09  Jerome Lambourg  <lambourg@adacore.com>

* libgnarl/s-taprop__qnx.adb: Fix incorrect casing for pthread_self.
* tracebak.c: Add support for tracebacks in QNX.

2017-11-09  Eric Botcazou  <ebotcazou@adacore.com>

* exp_aggr.adb (Aggr_Size_OK): Bump base limit from 50000 to 500000.

2017-11-09  Yannick Moy  <moy@adacore.com>

* erroutc.adb, set_targ.adb: Remove pragma Annotate for CodePeer
justification.

2017-11-09  Joel Brobecker  <brobecker@adacore.com>

* doc/gnat_ugn/platform_specific_information.rst: Document packages
needed on GNU/Linux by GNAT.
* gnat_ugn.texi: Regenerate.

2017-11-09  Hristian Kirtchev  <kirtchev@adacore.com>

* contracts.adb (Analyze_Contracts): Remove the three parameter
version. This routine now only analyzes contracts and does not perform
any freezing actions.
(Analyze_Previous_Contracts): Removed.
(Freeze_Previous_Contracts): New routine.
* contracts.ads (Analyze_Previous_Contracts): Removed.
(Freeze_Previous_Contracts): New routine.
* sem_ch3.adb (Analyze_Declarations): Analyze the contract of an
enclosing package spec regardless of whether the list denotes the
visible or private declarations.  Fix the removal of partial state
refinements when the context is a package spec.
* sem_ch6.adb (Analyze_Subprogram_Body_Helper): Freeze previous
contracts.
* sem_ch7.adb (Analyze_Package_Body_Helper): Freeze previous contracts.
* sem_ch9.adb (Analyze_Entry_Body): Freeze previous contracts.
(Analyze_Protected_Body): Freeze previous contracts.
(Analyze_Task_Body): Freeze previous contracts.
* sem_prag.adb: Comment reformatting.

2017-11-09  Bob Duff  <duff@adacore.com>

* libgnarl/g-thread.ads, libgnarl/g-thread.adb: (Make_Independent):
Export this so users can use it without importing
System.Tasking.Utilities.
* libgnarl/s-tassta.adb (Vulnerable_Complete_Task): Relax assertion
that fails when Make_Independent is called on a user task.
* libgnarl/s-taskin.ads (Master_Of_Task): Avoid unusual
capitalization style ((style) bad casing of "Master_of_Task").

From-SVN: r254580

gimple-ssa-store-merging.c (count_multiple_uses): New function.

* gimple-ssa-store-merging.c (count_multiple_uses): New function.
(split_group): Add total_orig and total_new arguments, estimate the
number of statements related to the store group without store merging
and with store merging.
(imm_store_chain_info::output_merged_store): Adjust split_group
callers, punt if estimated number of statements with store merging
is not smaller than estimated number of statements without it.
Formatting fix.
(handled_load): Remove has_single_use checks.
(pass_store_merging::process_store): Likewise.

From-SVN: r254579

[multiple changes]

2017-11-09  Ed Schonberg  <schonberg@adacore.com>

* sem_ch12.adb (Analyze_Subprogram_Instantiation): Correct use of
uninitialized variable uncovered by Codepeer.

2017-11-09  Arnaud Charlet  <charlet@adacore.com>

* namet.adb: Replace pragma Assume by pragma Assert to fix bootstrap.

2017-11-09  Javier Miranda  <miranda@adacore.com>

* doc/gnat_rm/standard_and_implementation_defined_restrictions.rst:
(Static_Dispatch_Tables): Minor rewording.
* gnat_rm.texi: Regenerate.

2017-11-09  Justin Squirek  <squirek@adacore.com>

* sem_ch8.adb (Analyze_Use_Package): Remove forced installation of
use_clauses within instances.
(Use_One_Package): Add condition to check for "hidden" open scopes to
avoid skipping over packages that have not been properly installed even
though they are visible.

2017-11-09  Ed Schonberg  <schonberg@adacore.com>

* sem_ch4.adb (Analyze_Selected_Component): Reject properly a call to a
private operation of a protected type, when the type has no visible
operations.

From-SVN: r254578

[multiple changes]

2017-11-09  Javier Miranda  <miranda@adacore.com>

* rtsfind.ads (RE_Id, RE_Unit_Table): Add RE_HT_Link.
* exp_disp.adb (Make_DT): Initialize the HT_Link field of the TSD only
if available.

2017-11-09  Bob Duff  <duff@adacore.com>

* exp_ch4.adb, exp_ch9.adb, exp_prag.adb, par-ch3.adb, sem_aggr.adb,
sem_ch12.adb, sem_ch13.adb, sem_ch4.adb, sem_disp.adb, sem_prag.adb,
sem_res.adb, sem_util.adb: Get rid of warnings about uninitialized
variables.

From-SVN: r254577

exp_disp.adb (Make_DT): Default initialize Ifaces_List and Ifaces_Comp_List.

2017-11-09  Yannick Moy  <moy@adacore.com>

* exp_disp.adb (Make_DT): Default initialize Ifaces_List and
Ifaces_Comp_List.

From-SVN: r254576

re PR tree-optimization/82902 (ICE verify_ssa failed during GIMPLE pass: phiprop)

2017-11-09  Richard Biener  <rguenther@suse.de>

PR tree-optimization/82902
* tree-ssa-phiprop.c (propagate_with_phi): Test proper type.

* g++.dg/torture/pr82902.C: New testcase.

From-SVN: r254575

Initialize split_branch_probability (PR target/82863).

2017-11-09  Martin Liska  <mliska@suse.cz>

PR target/82863
* emit-rtl.c (init_emit_regs): Initialize split_branch_probability to
uninitialized.
2017-11-09  Martin Liska  <mliska@suse.cz>

PR target/82863
* gcc.dg/pr82863.c: New test.

From-SVN: r254574

[multiple changes]

2017-11-09  Pascal Obry  <obry@adacore.com>

* libgnarl/s-taprop__mingw.adb: On Windows, initialize the thead handle
only for foreign threads.  We initialize the thread handle only if not
yet initialized. This happens in Enter_Task for foreign threads only.
But for native threads (Ada tasking) we do want to keep the real
handle (from Create_Task) to be able to free the corresponding
resources in Finalize_TCB (CloseHandle).

2017-11-09  Yannick Moy  <moy@adacore.com>

* sem_attr.adb (Analyze_Attribute): Default initialize P_Type,
P_Base_Type.
(Error_Attr_P): Fix name in pragma No_Return.
(Unexpected_Argument): Add pragma No_Return.
(Placement_Error): Add pragma No_Return.

2017-11-09  Javier Miranda  <miranda@adacore.com>

* exp_disp.adb (Elab_Flag_Needed): Elaboration flag not needed when the
dispatch table is statically built.
(Make_DT): Declare constant the Interface_Table object associated with
an statically built dispatch table. For this purpose the Offset_To_Top
value of each interface is computed using the dummy object.
* exp_ch3.adb (Build_Init_Procedure): Do not generate code initializing
the Offset_To_Top field of secondary dispatch tables when the dispatch
table is statically built.
(Initialize_Tag): Do not generate calls to Register_Interface_Offset
when the dispatch table is statically built.
* doc/gnat_rm/standard_and_implementation_defined_restrictions.rst:
Document the new GNAT restriction Static_Dispatch_Tables.
* gnat_rm.texi: Regenerate.

2017-11-09  Hristian Kirtchev  <kirtchev@adacore.com>

* sem_aggr.adb (Resolve_Delta_Record_Aggregate): Reorder declarations
to avoid a dormant bug.

2017-11-09  Jerome Lambourg  <lambourg@adacore.com>

* init.c: Define missing __gnat_alternate_stack for QNX. Set it to 0,
as such capability is not available on the OS.
* link.c: Make sure linker options for QNX are correct.
* libgnarl/s-osinte__qnx.ads: Add some missing bindings to pthread.
* libgnarl/s-taprop__qnx.adb: New, derived from s-taprop__posix.adb. This brings
in particular a workaround with locks priority ceiling where a higher
priority task is allowed to lock a lower ceiling priority lock. This
also fixes the scheduling of FIFO tasks when the priority of a task is
lowered.
* libgnat/system-qnx-aarch64.ads: Fix priority ranges.

2017-11-09  Yannick Moy  <moy@adacore.com>

* erroutc.adb (Output_Error_Msgs): Justify CodePeer false positive
message.
* gnatbind.adb (Scan_Bind_Arg): Simplify test to remove always true
condition.
* namet.adb (Copy_One_Character): Add assumption for static analysis,
as knowledge that Hex(2) is in the range 0..255 is too complex for
CodePeer.
(Finalize): Add assumption for static analysis, as the fact that there
are symbols in the table depends on a global invariant at this point in
the program.
* set_targ.adb (Check_Spaces): Justify CodePeer false positive message.
* stylesw.adb (Save_Style_Check_Options): Rewrite to avoid test always
true.

From-SVN: r254573

[multiple changes]

2017-11-09  Javier Miranda  <miranda@adacore.com>

* libgnat/s-rident.ads (Static_Dispatch_Tables): New restriction name.
* exp_disp.adb (Building_Static_DT): Check restriction.
(Building_Static_Secondary_DT): Check restriction.
(Make_DT): Initialize the HT_Link to No_Tag.
* opt.ads (Static_Dispatch_Tables): Rename flag...
(Building_Static_Dispatch_Tables): ... into this.  This will avoid
conflict with the restriction name.
* gnat1drv.adb: Update.
* exp_aggr.adb (Is_Static_Dispatch_Table_Aggregate): Update.
* exp_ch3.adb (Expand_N_Object_Declaration): Update.

2017-11-09  Pascal Obry  <obry@adacore.com>

* libgnarl/s-taprop__mingw.adb: Minor code clean-up.  Better using a
named number.

From-SVN: r254572

exp_ch3.adb, [...]: Minor reformatting.

gcc/ada/

2017-11-09  Ed Schonberg  <schonberg@adacore.com>

* exp_ch3.adb, gnat1drv.adb, namet.adb, namet.ads, sem_aggr.adb,
sem_ch2.adb, sem_ch4.adb: Minor reformatting.
* sem_res.adb (Resolve_Entity_Name): Suppress spurious error on read of
out parameter when in Ada_83 mode, the oarameter is of a composite
type, and it appears as the prefix of an attribute.

2017-11-09  Bob Duff  <duff@adacore.com>

* sinfo.ads: Minor comment fix.

2017-11-09  Hristian Kirtchev  <kirtchev@adacore.com>

* sem_prag.ads: Add pragmas Unmodified and Unreferenced to table
Pragma_Significant_In_SPARK.

gcc/testsuite/

2017-11-09  Hristian Kirtchev  <kirtchev@adacore.com>

* gnat.dg/unreferenced.adb: New testcase.

2017-11-09  Ed Schonberg  <schonberg@adacore.com>

* gnat.dg/out_param.adb: New testcase.

From-SVN: r254571

[multiple changes]

2017-11-09  Yannick Moy  <moy@adacore.com>

* binde.adb (Diagnose_Elaboration_Problem): Mark procedure No_Return.
* checks.adb (Apply_Scalar_Range_Check): Rescope variable OK closer to
use.  Default initialize Hi, Lo.
(Selected_Range_Checks): Retype Num_Checks more precisely.
(Determine_Range, Determine_Range_R): Default initialize Hi_Right,
Lo_Right.
* contracts.adb (Process_Contract_Cases): Mark parameter Stmts as
Unmodified.
(Process_Postconditions): Mark parameter Stmts as Unmodified.
* exp_attr.adb (Expand_Loop_Entry_Attribute): Default initialize Blk.
* exp_ch4.adb (Expand_N_Allocator): Default initialize Typ.
(Expand_Concatenate): Default initialize High_Bound.
(Optimize_Length_Comparison): Default initialize Ent, Index.
* exp_ch5.adb (Expand_Predicated_Loop): Default initialize L_Hi and
L_Lo.
* exp_ch6.adb (Expand_N_Extended_Return_Statement): Default initialize
Return_Stmt.
* exp_ch9.adb (Expand_Entry_Barrier): Default initialize Func_Body and
remove pragma Warnings(Off).
* exp_imgv.adb (Expand_Image_Attribute): Default initialize Tent.
* exp_util.adb (Find_Interface_Tag): Default initialize AI_Tag.
* freeze.adb (Check_Component_Storage_Order): Default initialize
Comp_Byte_Aligned rather than silencing messages with pragma
Warnings(Off), which does not work for CodePeer initialization
messages, and given that here the possible read of an unitialized value
depends on a proper use of parameters by the caller.
* inline.adb (Expand_Inlined_Call): Default initialize Lab_Decl, Targ.
* sem_ch12.adb (Build_Operator_Wrapper): Default initialize Expr.
* sem_ch3.adb (Build_Derived_Array_Type): Default initialize
Implicit_Base.
* sem_ch4.adb (List_Operand_Interps): Default initialize Nam and remove
pragma Warnings(Off).
(Analyze_Case_Expression): Rescope checking block within branch where
Others_Present is set by the call to Check_Choices.
* sem_ch5.adb (Analyze_Assignment): Default initialize
Save_Full_Analysis.
* sem_ch6.adb (Analyze_Function_Return): Default initialize Obj_Decl,
and restructure code to defend against previous errors, so that, in
that case, control does not flow to the elsif condition which read an
uninitialized Obj_Decl.
* sem_ch9.adb (Analyze_Requeue): Default initialize Synch_Type.
(Check_Interfaces): Default initialize Full_T_Ifaces and Priv_T_Ifaces,
which seem to be left uninitialized and possibly read in some cases.
* sem_dim.adb (Analyze_Aspect_Dimension_System): Retype Position more
precisely.  This requires to exchange the test for exiting in case of
too many positions and the increment to Position, inside the loop.
* sem_eval.adb (Eval_Concatenation): Default initialize Folded_Val,
which cannot be read uninitialized, but the reasons for that are quite
subtle.
* sem_intr.adb (Check_Intrinsic_Call): Default initialize Rtyp.
* sem_prag.adb (Collect_Subprogram_Inputs_Outputs): Default initialize
Spec_Id.
* sem_res.adb (Make_Call_Into_Operator): Default initialize Opnd_Type,
and test for presence of non-null Opnd_Type before testing its scope,
in a test which would read its value uninitialized, and is very rarely
exercized (it depends on the presence of an extension of System).
* sem_spark.ads: Update comment to fix name of main analysis procedure.
* sem_warn.adb (Warn_On_Known_Condition): Default initialize
Test_Result.
* set_targ.adb (FailN): Mark procedure with No_Return.
* stylesw.adb (Save_Style_Check_Options): Delete useless code to
initialize all array Options to white space, as there is already code
doing the same for the remaining positions in Options at the end of the
procedure.

2017-11-09  Eric Botcazou  <ebotcazou@adacore.com>

* exp_ch11.adb (Possible_Local_Raise): Do not issue the warning for
generic instantiations either.

From-SVN: r254570

[multiple changes]

2017-11-09  Piotr Trojanek  <trojanek@adacore.com>

* sem_prag.adb (Analyze_Part_Of): Reword error message.
(Get_SPARK_Mode_Type): Do not raise Program_Error in case pragma
SPARK_Mode appears with an illegal mode, treat this as a non-existent
mode.

2017-11-09  Ed Schonberg  <schonberg@adacore.com>

* sem_ch4.adb (Analyze_Call): Reject a call to a function that returns
a limited view of a type T declared in unit U1, when the function is
declared in another unit U2 and the call appears in a procedure within
another unit.

2017-11-09  Justin Squirek  <squirek@adacore.com>

* sem_ch8.adb (Analyze_Use_Package): Force installation of use_clauses
when processing generic instances.

2017-11-09  Bob Duff  <duff@adacore.com>

* namet.ads, namet.adb (Valid_Name_Id): New subtype that excludes
Error_Name and No_Name.  Use this (versus Name_Id) to indicate which
objects can have those special values. Valid_Name_Id could usefully be
used all over the compiler front end, but that's too much trouble for
now. If we did that, we might want to rename:
Name_Id --> Optional_Name_Id, Valid_Name_Id --> Name_Id.
For parameters of type Valid_Name_Id, remove some redundant tests,
including the ones found by CodePeer.  Use Is_Valid_Name instead of
membership test when appropriate.
(Error_Name_Or_No_Name): Delete this; it's no longer needed.
* sem_ch2.adb (Analyze_Identifier): Use "not Is_Valid_Name" instead of
"in Error_Name_Or_No_Name".
(Check_Parameterless_Call): Use "not Is_Valid_Name" instead of "in
Error_Name_Or_No_Name".

From-SVN: r254569

gnat1drv.adb (Adjust_Global_Switches): Suppress warnings in codepeer mode here unless -gnateC is specified.

gcc/ada/

2017-11-09  Arnaud Charlet  <charlet@adacore.com>

* gnat1drv.adb (Adjust_Global_Switches): Suppress warnings in codepeer
mode here unless -gnateC is specified.
* switch-c.adb (Scan_Front_End_Switches): Do not suppress warnings with
-gnatC here.

2017-11-09  Piotr Trojanek  <trojanek@adacore.com>

* lib-writ.adb (Write_ALI): Remove processing of the frontend xrefs as
part of the ALI writing; they are now processed directly from memory
when requested by the backend.
* lib-xref.ads (Collect_SPARK_Xrefs): Remove.
(Iterate_SPARK_Xrefs): New routine for iterating over frontend xrefs.
* lib-xref-spark_specific.adb (Traverse_Compilation_Unit): Remove.
(Add_SPARK_File): Remove.
(Add_SPARK_Xref): Refactored from removed code; filters xref entries
that are trivially uninteresting to the SPARK backend.
* spark_xrefs.ads: Remove code that is no longer needed.
* spark_xrefs.adb (dspark): Adapt to use Iterate_SPARK_Xrefs.

2017-11-09  Hristian Kirtchev  <kirtchev@adacore.com>

* sem_elab.adb: Update the documentation on adding a new elaboration
schenario. Add new hash table Recorded_Top_Level_Scenarios.
(Is_Check_Emitting_Scenario): Removed.
(Is_Recorded_Top_Level_Scenario): New routine.
(Kill_Elaboration_Scenario): Reimplemented.
(Record_Elaboration_Scenario): Mark the scenario as recorded.
(Set_Is_Recorded_Top_Level_Scenario): New routine.
(Update_Elaboration_Scenario): Reimplemented.
* sinfo.adb (Is_Recorded_Scenario): Removed.
(Set_Is_Recorded_Scenario): Removed.
* sinfo.ads: Remove attribute Is_Recorded_Scenario along with
occurrences in nodes.
(Is_Recorded_Scenario): Removed along with pragma Inline.
(Set_Is_Recorded_Scenario): Removed along with pragma Inline.

2017-11-09  Piotr Trojanek  <trojanek@adacore.com>

* sem_prag.adb (Analyze_Part_Of): Change "designate" to "denote" in
error message.

2017-11-09  Justin Squirek  <squirek@adacore.com>

* sem_res.adb (Resolve_Allocator): Add warning messages corresponding
to the allocation of an anonymous access-to-controlled object.

gcc/testsuite/

2017-11-09  Hristian Kirtchev  <kirtchev@adacore.com>

* gnat.dg/elab3.adb, gnat.dg/elab3.ads, gnat.dg/elab3_pkg.adb,
gnat.dg/elab3_pkg.ads: New testcase.

2017-11-09  Pierre-Marie de Rodat  <derodat@adacore.com>

    * gnat.dg/controlled2.adb, gnat.dg/controlled4.adb, gnat.dg/finalized.adb:
    Disable the new warning from GNAT.

From-SVN: r254568

Remove non needed check in bmp_iter_set_init (PR tree-optimization/82669).

2017-11-09  Martin Liska  <mliska@suse.cz>

PR tree-optimization/82669
* sbitmap.h (bmp_iter_set_init): Remove non needed check.

From-SVN: r254567

[multiple changes]

2017-11-09  Jerome Lambourg  <lambourg@adacore.com>

* sigtramp-qnx.c: Fix obvious typo.

2017-11-09  Doug Rupp  <rupp@adacore.com>

* libgnarl/s-taprop__linux.adb (Monotonic_Clock): Minor reformatting.

2017-11-09  Ed Schonberg  <schonberg@adacore.com>

* sem_res.adb (Resolve): If expression is an entity whose type has
implicit dereference, generate reference to it, because no reference is
generated for an overloaded entity during analysis, given that its
identity may not be known.

2017-11-09  Javier Miranda  <miranda@adacore.com>

* exp_disp.adb (Expand_Interface_Thunk): Replace substraction of
offset-to-top field by addition.
(Make_Secondary_DT): Initialize the offset-to-top field with a negative
offset.
* exp_ch3.adb (Build_Offset_To_Top_Function): Build functions that
return a negative offset-to-top value.
(Initialize_Tag): Invoke runtime services Set_Dynamic_Offset_To_Top and
Set_Static_Offset_To_Top passing a negative offet-to-top value;
initialize also the offset-to-top field with a negative offset.
* libgnat/a-tags.adb (Base_Address): Displace the pointer by means of
an addition since the offset-to-top field is now a negative value.
(Displace): Displace the pointer to the object means of a substraction
since it is now a negative value.
(Set_Dynamic_Offset_to_top): Displace the pointer to the object by
means of a substraction since it is now a negative value.

2017-11-09  Eric Botcazou  <ebotcazou@adacore.com>

* gnat1drv.adb (Gnat1drv): Call Errout.Finalize (Last_Call => True)
before Errout.Output_Messages also in the case of compilation errors.

2017-11-09  Javier Miranda  <miranda@adacore.com>

* doc/gnat_ugn/the_gnat_compilation_model.rst (Interfacing with C++ at
the Class Level): Fix error interfacing with C strings.
* gnat_ugn.texi: Regenerate.

2017-11-09  Jerome Lambourg  <lambourg@adacore.com>

* system-qnx-aarch64.ads: Fix the priority constants.
* s-osinte__qnx.ads: Fix constants for handling the locking protocols
and scheduling.
* s-osinte__qnx.adb: New file , prevents the use of priority 0 that
corresponds to an idle priority on QNX.

2017-11-09  Piotr Trojanek  <trojanek@adacore.com>

* sem_prag.adb, sem_util.adb, sem_elab.adb: Fix minor typos in
comments.

From-SVN: r254566

[multiple changes]

2017-11-09  Piotr Trojanek  <trojanek@adacore.com>

* lib-xref-spark_specific.adb (Add_SPARK_Xrefs): Ignore loop parameters
in expression funtions that are expanded into variables.

2017-11-09  Piotr Trojanek  <trojanek@adacore.com>

* sem_util.adb: Minor whitespace cleanup.

2017-11-09  Jerome Lambourg  <lambourg@adacore.com>

* libgnarl/s-taprop__qnx.adb: Refine aarch64-qnx. Use the POSIX
s-taprop version rather than a custom one.
* sigtramp-qnx.c (aarch64-qnx): Implement the signal trampoline.

From-SVN: r254563

GCOV: support multiple functions per a line (PR gcov-profile/48463)

2017-11-09  Martin Liska  <mliska@suse.cz>

PR gcov-profile/48463
* coverage.c (coverage_begin_function): Output also end locus
of a function and information whether the function is
artificial.
* gcov-dump.c (tag_function): Parse and print the information.
* gcov.c (INCLUDE_MAP): Add include.
(INCLUDE_SET): Likewise.
(struct line_info): Move earlier in the source file because
of vector<line_info> in function_info structure.
(line_info::line_info): Likewise.
(line_info::has_block): Likewise.
(struct source_info): Add new member index.
(source_info::get_functions_at_location): New function.
(function_info::group_line_p): New function.
(output_intermediate_line): New function.
(output_intermediate_file): Use the mentioned function.
(struct function_start): New.
(struct function_start_pair_hash): Likewise.
(process_file): Add code that identifies group functions.
Assign lines either to global or function scope.
(generate_results): Skip artificial functions.
(find_source): Assign index for each source file.
(read_graph_file): Read new flag artificial and end_line.
(add_line_counts): Assign it either to global of function scope.
(accumulate_line_counts): Isolate core of the function to
accumulate_line_info and call it for both function and global
scope lines.
(accumulate_line_info): New function.
(output_line_beginning): Fix GNU coding style.
(print_source_line): New function.
(output_line_details): Likewise.
(output_function_details): Likewise.
(output_lines): Iterate both source (global) scope and function
scope.
(struct function_line_start_cmp): New class.
* doc/gcov.texi: Reflect changes in documentation.

From-SVN: r254562

re PR debug/82837 (ICE in output_operand: invalid expression as operand)

PR debug/82837
* dwarf2out.c (const_ok_for_output_1): Reject NEG in addition to NOT.
(mem_loc_descriptor): Handle (const (neg (...))) as (neg (const (...)))
and similarly for not instead of neg.

* gcc.dg/debug/dwarf2/pr82837.c: New test.

From-SVN: r254561

Add option to force indirect calls for x86

This patch adds a -mforce-indirect-call option to force all calls
or tail calls on x86_64 between functions to indirect. This is similar to the
large code model, but doesn't affect jumps inside functions, so has much
less run time overhead.

This is useful with Intel Processor Trace (PT). PT has precise timing
for indirect calls/jumps, but not for direct ones. So if we can force
them to indirect it allows to time every function relatively accurately
(minus the overhead of the indirect branch)

Without this short functions often don't see a timing update and cannot
be measured.

The timing requires at least Skylake or Goldmont based CPUs.

I made it an option. Originally I tried to make it a new code model,
but since it can be combined with other code models (medium, pic, kernel
etc.) this turned out to be too many combinations.

For example with gcc. This first column is a ns time stamp for the
functions.

$ perf record -e intel_pt/noretcomp=1,cyc=1,cyc_thresh=1/u ./cc1 -O3 hello.c
$ perf script --itrace=cr -F callindent,time,sym,addr --ns  | sed -n 180000,182000p | less
...
1184596.432756920:                             build_int_cst                        =>           79c9de c_common_nodes_and_builtins
1184596.432756921:                             tree_cons                            =>           ee2080 tree_cons
1184596.432756938:                                 ggc_internal_alloc               =>           80f3e0 ggc_internal_alloc
1184596.432756951:                                     memset@plt                   =>           598af0 memset@plt
1184596.432756967:                                     __memset_avx2_unaligned_erms =>           80f605 ggc_internal_alloc
1184596.432756969:                                 ggc_internal_alloc               =>           ee20a2 tree_cons
1184596.432756973:                             tree_cons                            =>           79c9f4 c_common_nodes_and_builtins
1184596.432756974:                             build_int_cst                        =>           ef9a40 build_int_cst
1184596.432756996:                                 wide_int_to_tree                 =>           ef93a0 wide_int_to_tree
1184596.432757000:                                     wi::force_to_size            =>           f48f70 wi::force_to_size
1184596.432757005:                                     canonize                     =>           ef94de wide_int_to_tree
1184596.432757021:                                     get_int_cst_ext_nunits       =>           ee1960 get_int_cst_ext_nunits
1184596.432757026:                                     get_int_cst_ext_nunits       =>           ef94fe wide_int_to_tree
1184596.432757042:                                     tree_int_cst_elt_check       =>           83e310 tree_int_cst_elt_check
1184596.432757044:                                     tree_int_cst_elt_check       =>           ef9761 wide_int_to_tree
1184596.432757046:                                 wide_int_to_tree                 =>           ef9a9b build_int_cst

gcc/:
2017-11-08  Andi Kleen  <ak@linux.intel.com>

* config/i386/i386.opt: Add -mforce-indirect-call.
* config/i386/predicates.md: Check for flag_force_indirect_call.
* doc/invoke.texi: Document -mforce-indirect-call

gcc/testsuite/:
2017-11-08  Andi Kleen  <ak@linux.intel.com>

* gcc.target/i386/force-indirect-call-1.c: New test.
* gcc.target/i386/force-indirect-call-2.c: New test.
* gcc.target/i386/force-indirect-call-3.c: New test.

From-SVN: r254560

Daily bump.

From-SVN: r254559

re PR fortran/82841 (Segfault in gfc_simplify_transfer)

2017-11-08  Steven G. Kargl  <kargl@kgcc.gnu.org>

PR Fortran/82841
* simplify.c(gfc_simplify_transfer): Do not dereference a NULL pointer.
Unwrap a short line.

2017-11-08  Steven G. Kargl  <kargl@kgcc.gnu.org>

PR Fortran/82841
* gfortran.dg/transfer_simplify_11.f90: new test.

From-SVN: r254555

RISC-V: Fix build error

  - This build error was indroduced by "RISC-V: Implement movmemsi"
    and "RISC-V: Support -mpreferred-stack-boundary flag"

gcc/ChangeLog

2017-11-08  Kito Cheng  <kito.cheng@gmail.com>

        * config/riscv/riscv-protos.h (riscv_slow_unaligned_access_p):
        New extern.
        (MOVE_RATIO): Use riscv_slow_unaligned_access_p.
        config/riscv/riscv.c (predict.h): New include.
        (riscv_slow_unaligned_access_p): No longer static.
        (riscv_block_move_straight): Add require.
        config/riscv/riscv-protos.h (riscv_hard_regno_nregs): Delete.

From-SVN: r254554

re PR fortran/82884 (ICE in gfc_resolve_character_array_constructor, at fortran/array.c:2069)

2017-11-08  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/82884
* arith.c (gfc_hollerith2character): Clear pad.

2017-11-08  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/82884
* gfortran.dg/hollerith_character_array_constructor.f90: New test.

From-SVN: r254553

re PR target/82855 (AVX512: replace OP+movemask with OP_mask+ktest)

PR target/82855
* config/i386/sse.md (<avx512>_eq<mode>3<mask_scalar_merge_name>,
<avx512>_eq<mode>3<mask_scalar_merge_name>_1): Use
nonimmediate_operand predicate for operand 1 instead of
register_operand.

From-SVN: r254552

[AArch64] Add STP pattern to store a vec_concat of two 64-bit registers

On top of the previous vec_merge simplifications [1] we can add this pattern to perform
a store of a vec_concat of two 64-bit values in distinct registers as an STP.
This avoids constructing such a vector explicitly in a register and storing it as
a Q register.
This way for the code in the testcase we can generate:

construct_lane_1:
        ldp     d1, d0, [x0]
        fmov    d3, 1.0e+0
        fmov    d2, 2.0e+0
        fadd    d4, d1, d3
        fadd    d5, d0, d2
        stp     d4, d5, [x1, 32]
        ret

construct_lane_2:
        ldp     x2, x0, [x0]
        add     x3, x2, 1
        add     x4, x0, 2
        stp     x3, x4, [x1, 32]
        ret

instead of the current:
construct_lane_1:
        ldp     d0, d1, [x0]
        fmov    d3, 1.0e+0
        fmov    d2, 2.0e+0
        fadd    d0, d0, d3
        fadd    d1, d1, d2
        dup     v0.2d, v0.d[0]
        ins     v0.d[1], v1.d[0]
        str     q0, [x1, 32]
        ret

construct_lane_2:
        ldp     x2, x3, [x0]
        add     x0, x2, 1
        add     x2, x3, 2
        dup     v0.2d, x0
        ins     v0.d[1], x2
        str     q0, [x1, 32]
        ret

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

[1] https://gcc.gnu.org/ml/gcc-patches/2017-06/msg00272.html
    https://gcc.gnu.org/ml/gcc-patches/2017-06/msg00273.html
    https://gcc.gnu.org/ml/gcc-patches/2017-06/msg00274.html

* config/aarch64/aarch64-simd.md (store_pair_lanes<mode>):
New pattern.
* config/aarch64/constraints.md (Uml): New constraint.
* config/aarch64/predicates.md (aarch64_mem_pair_lanes_operand): New
predicate.

* gcc.target/aarch64/store_v2vec_lanes.c: New test.

From-SVN: r254551

[simplify-rtx] Simplify vec_merge of vec_duplicates into vec_concat

Another vec_merge simplification that's missing from simplify-rtx.c is transforming
a vec_merge of two vec_duplicates. For example:
(set (reg:V2DF 80)
    (vec_merge:V2DF (vec_duplicate:V2DF (reg:DF 84))
        (vec_duplicate:V2DF (reg:DF 81))
        (const_int 2)))

Can be transformed into the simpler:
(set (reg:V2DF 80)
    (vec_concat:V2DF (reg:DF 81)
                (reg:DF 84)))

I believe this should always be beneficial.
I'm still looking into finding a small testcase demonstrating this, but on aarch64 SPEC
I've seen this eliminate some really bizzare codegen where GCC was generating nonsense like:
  ldr q18, [sp, 448]
  ins v18.d[0], v23.d[0]
  ins v18.d[1], v22.d[0]

With q18 being pushed and popped off the stack in the prologue and epilogue of the function!
These are large files from SPEC that I haven't been able to analyse yet as to why GCC even attempts
to do that, but with this patch it doesn't try to load a register and overwrite all its lanes.
This patch shaves off about 5k of code size from zeusmp on aarch64 at -O3, so I believe it's a good
thing to do.

* simplify-rtx.c (simplify_ternary_operation): Simplify vec_merge
of two vec_duplicates into a vec_concat.

From-SVN: r254550

vec_merge + vec_duplicate + vec_concat simplification

Another vec_merge simplification that's missing is transforming:
(vec_merge (vec_duplicate x) (vec_concat (y) (z)) (const_int N))
into
(vec_concat x z) if N == 1 (0b01) or
(vec_concat y x) if N == 2 (0b10)

For the testcase in this patch on aarch64 this allows us to try matching during combine the pattern:
(set (reg:V2DI 78 [ x ])
    (vec_concat:V2DI
        (mem:DI (reg/v/f:DI 76 [ y ]) [1 *y_4(D)+0 S8 A64])
        (mem:DI (plus:DI (reg/v/f:DI 76 [ y ])
                (const_int 8 [0x8])) [1 MEM[(long long int *)y_4(D) + 8B]+0 S8 A64])))

rather than the more complex:
(set (reg:V2DI 78 [ x ])
    (vec_merge:V2DI (vec_duplicate:V2DI (mem:DI (plus:DI (reg/v/f:DI 76 [ y ])
                    (const_int 8 [0x8])) [1 MEM[(long long int *)y_4(D) + 8B]+0 S8 A64]))
        (vec_duplicate:V2DI (mem:DI (reg/v/f:DI 76 [ y ]) [1 *y_4(D)+0 S8 A64]))
        (const_int 2 [0x2])))

We don't actually have an aarch64 pattern for the simplified version above, but it's a simple enough
form to add, so this patch adds such a pattern that performs a concatenated load of two 64-bit vectors
in adjacent memory locations as a single Q-register LDR. The new aarch64 pattern is needed to demonstrate
the effectiveness of the simplify-rtx change, so I've kept them together as one patch.

Now for the testcase in the patch we can generate:
construct_lanedi:
        ldr     q0, [x0]
        ret

construct_lanedf:
        ldr     q0, [x0]
        ret

instead of:
construct_lanedi:
        ld1r    {v0.2d}, [x0]
        ldr     x0, [x0, 8]
        ins     v0.d[1], x0
        ret

construct_lanedf:
        ld1r    {v0.2d}, [x0]
        ldr     d1, [x0, 8]
        ins     v0.d[1], v1.d[0]
        ret

The new memory constraint Utq is needed because we need to allow only the Q-register addressing modes but
the MEM expressions in the RTL pattern have 64-bit vector modes, and if we don't constrain them they will
allow the D-register addressing modes during register allocation/address mode selection, which will produce
invalid assembly.

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

* simplify-rtx.c (simplify_ternary_operation, VEC_MERGE):
Simplify vec_merge of vec_duplicate and vec_concat.
* config/aarch64/constraints.md (Utq): New constraint.
* config/aarch64/aarch64-simd.md (load_pair_lanes<mode>): New
define_insn.

* gcc.target/aarch64/load_v2vec_lanes_1.c: New test.

From-SVN: r254549

Simplify vec_merge of vec_duplicate with const_vector

I'm trying to improve some of the RTL-level handling of vector lane operations on aarch64 and that
involves dealing with a lot of vec_merge operations. One simplification that I noticed missing
from simplify-rtx are combinations of vec_merge with vec_duplicate.
In this particular case:
(vec_merge (vec_duplicate (X)) (const_vector [A, B]) (const_int N))

which can be replaced with

(vec_concat (X) (B)) if N == 1 (0b01) or
(vec_concat (A) (X)) if N == 2 (0b10).

For the aarch64 testcase in this patch this simplifications allows us to try to combine:
(set (reg:V2DI 77 [ x ])
    (vec_concat:V2DI (mem:DI (reg:DI 0 x0 [ y ]) [1 *y_3(D)+0 S8 A64])
        (const_int 0 [0])))

instead of the more complex:
(set (reg:V2DI 77 [ x ])
    (vec_merge:V2DI (vec_duplicate:V2DI (mem:DI (reg:DI 0 x0 [ y ]) [1 *y_3(D)+0 S8 A64]))
        (const_vector:V2DI [
                (const_int 0 [0])
                (const_int 0 [0])
            ])
        (const_int 1 [0x1])))

For the simplified form above we already have an aarch64 pattern: *aarch64_combinez<mode> which
is missing a DI/DFmode version due to an oversight, so this patch extends that pattern as well to
use the VDC mode iterator that includes DI and DFmode (as well as V2HF which VD_BHSI was missing).
The aarch64 hunk is needed to see the benefit of the simplify-rtx.c hunk, so I didn't split them
into separate patches.

Before this for the testcase we'd generate:
construct_lanedi:
        movi    v0.4s, 0
        ldr     x0, [x0]
        ins     v0.d[0], x0
        ret

construct_lanedf:
        movi    v0.2d, 0
        ldr     d1, [x0]
        ins     v0.d[0], v1.d[0]
        ret

but now we can generate:
construct_lanedi:
        ldr     d0, [x0]
        ret

construct_lanedf:
        ldr     d0, [x0]
        ret

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

* simplify-rtx.c (simplify_ternary_operation, VEC_MERGE):
Simplify vec_merge of vec_duplicate and const_vector.
* config/aarch64/predicates.md (aarch64_simd_or_scalar_imm_zero):
New predicate.
* config/aarch64/aarch64-simd.md (*aarch64_combinez<mode>): Use VDC
mode iterator.  Update predicate on operand 1 to
handle non-const_vec constants.  Delete constraints.
(*aarch64_combinez_be<mode>): Likewise for operand 2.

* gcc.target/aarch64/construct_lane_zero_1.c: New test.

From-SVN: r254548

[multiple changes]

2017-11-08  Piotr Trojanek  <trojanek@adacore.com>

* lib-xref.ads, lib-xref-spark_specific.adb
(Traverse_Compilation_Unit): Move declaration to package body.

2017-11-08  Hristian Kirtchev  <kirtchev@adacore.com>

* exp_spark.adb (Expand_SPARK_N_Object_Renaming_Declaration): Obtain
the type of the renaming from its defining entity, rather then the
subtype mark as there may not be a subtype mark.

2017-11-08  Jerome Lambourg  <lambourg@adacore.com>

* adaint.c, s-oscons-tmplt.c, init.c, libgnat/system-qnx-aarch64.ads,
libgnarl/a-intnam__qnx.ads, libgnarl/s-intman__qnx.adb,
libgnarl/s-osinte__qnx.ads, libgnarl/s-qnx.ads,
libgnarl/s-taprop__qnx.adb, s-oscons-tmplt.c, sigtramp-qnx.c,
terminals.c: Initial port of GNAT for aarch64-qnx

2017-11-08  Elisa Barboni  <barboni@adacore.com>

* exp_util.adb (Find_DIC_Type): Move...
* sem_util.ads, sem_util.adb (Find_DIC_Type): ... here.

2017-11-08  Justin Squirek  <squirek@adacore.com>

* sem_res.adb (Resolve_Allocator): Add info messages corresponding to
the owner and corresponding coextension.

2017-11-08  Ed Schonberg  <schonberg@adacore.com>

* sem_aggr.adb (Resolve_Delta_Aggregate): Divide into the
following separate procedures.
(Resolve_Delta_Array_Aggregate): Previous code form
Resolve_Delta_Aggregate.
(Resolve_Delta_Record_Aggregate): Extend previous code to cover latest
ARG decisions on the legality rules for delta aggregates for records:
in the case of a variant record, components from different variants
cannot be specified in the delta aggregate, and this must be checked
statically.

From-SVN: r254547

[multiple changes]

2017-11-08  Piotr Trojanek  <trojanek@adacore.com>

* spark_xrefs.ads (SPARK_Scope_Record): Remove File_Num component.
* lib-xref-spark_specific.adb (Add_SPARK_Scope): Skip initialization of
removed component.

2017-11-08  Gary Dismukes  <dismukes@adacore.com>

* sem_ch4.adb: Minor typo fix.

From-SVN: r254546