mmx.md (vec_extractv2si_0): Add (r,x) alternative.

* config/i386/mmx.md (vec_extractv2si_0): Add (r,x) alternative.
(*vec_extractv2si_0_zext_sse4): New insn pattern.
(*vec_extractv2si_0_zext): Ditto.
(*vec_extractv2si_1): Add (rm,x) alternative.
(*vec_extractv2si_1_zext): New insn pattern.
(*vec_extractv2si_zext_mem): Add "TARGET_MMX || TARGET_MMX_WITH_SSE"
insn constraint.

From-SVN: r273979

domwalk: use gcc_sort_r

2019-08-01  Richard Biener  <rguenther@suse.de>

* domwalk.c (bb_postorder): Remove static variable.
(cmp_bb_postorder): Adjust.
(sort_bbs_postorder): Adjust and use gcc_sort_r.
(dom_walker::walk): Adjust.

From-SVN: r273978

sort.cc: introduce gcc_sort_r

* sort.cc (sort_r_ctx): New struct.
(reorder23): Make templated on context type.
(reorder45): Ditto.
(cmp1): Ditto.  Adjust signature.
(netsort): Ditto.
(mergesort): Ditto.
[CHECKING_P] (cmp2to3): New static function.  Use it...
(gcc_qsort) [CHECKING_P]: ...here.
(gcc_sort_r): New function.
* system.h (sort_r_cmp_fn): New function typedef.
(qsort_chk): Adjust signature.
(gcc_sort_r): Declare.
* vec.c (qsort_chk_error): Adjust.
(qsort_chk): Adjust.

From-SVN: r273977

PR c++/90805 - detect narrowing in case values.

* decl.c (case_conversion): Detect narrowing in case values.

* c-c++-common/pr89888.c: Update expected dg-error.
* g++.dg/cpp0x/Wnarrowing17.C: New test.
* g++.dg/cpp0x/enum28.C: Update expected dg-error.

From-SVN: r273976

Implement C++20 p0202 - Add Constexpr Modifiers to Functions in <algorithm> and <utility> Headers.

2019-08-01  Edward Smith-Rowland  <3dw4rd@verizon.net>

Implement C++20 p0202 - Add Constexpr Modifiers to Functions
in <algorithm> and <utility> Headers.
Implement C++20 p1023 - constexpr comparison operators for std::array.
* include/bits/algorithmfwd.h (all_of, any_of, binary_search, copy,
copy_backward, copy_if, copy_n, equal_range, fill, find_end,
find_if_not, includes, is_heap, is_heap_until, is_partitioned,
is_permutation, is_sorted, is_sorted_until, iter_swap, lower_bound,
none_of, partition_copy, partition_point, remove, remove_if,
remove_copy, remove_copy_if, replace_copy, replace_copy_if,
reverse_copy, rotate_copy, uunique, upper_bound, adjacent_find, count,
count_if, equal, find, find_first_of, find_if, for_each, generate,
generate_n, lexicographical_compare, merge, mismatch, replace,
replace_if, search, search_n, set_difference, set_intersection,
set_symmetric_difference, set_union, transform, unique_copy):
Mark constexpr.
* include/bits/cpp_type_traits.h (__miter_base): Mark constexpr.
* include/bits/predefined_ops.h (_Iter_less_val::operator(),
_Val_less_iter::operator(), _Iter_equal_to_iter::operator(),
_Iter_equal_to_val::operator(), _Iter_equals_val::operator()):
 Use const ref instead of ref arg;
(_Iter_less_val, __iter_less_val, _Val_less_iter, __val_less_iter,
__iter_equal_to_iter, __iter_equal_to_val, __iter_comp_val,
_Iter_comp_val, _Val_comp_iter, __val_comp_iter, __iter_equals_val,
_Iter_equals_iter, __iter_comp_iter, _Iter_pred, __pred_iter,
_Iter_comp_to_val, __iter_comp_val, _Iter_comp_to_iter,
__iter_comp_iter): Mark constexpr.
* include/bits/stl_algo.h (__find_if, __find_if_not, __find_if_not_n,
__search, __search_n_aux, __search_n, __find_end, find_end, all_of,
none_of, any_of, find_if_not, is_partitioned, partition_point,
__remove_copy_if, remove_copy, remove_copy_if, copy_if, __copy_n,
copy_n, partition_copy, __remove_if, remove, remove_if, __adjacent_find,
__unique, unique, __unique_copy, reverse_copy, rotate_copy,
__unguarded_linear_insert, __insertion_sort, __unguarded_insertion_sort,
__final_insertion_sort, lower_bound, __upper_bound, upper_bound,
__equal_range, equal_range, binary_search, __includes, includes,
__next_permutation, __prev_permutation, __replace_copy_if, replace_copy,
replace_copy_if, __count_if, is_sorted, __is_sorted_until,
is_sorted_until, __is_permutation, is_permutation, for_each, find,
find_if, find_first_of, adjacent_find, count, count_if, search,
search_n, transform, replace, replace_if, generate, generate_n,
unique_copy, __merge, merge, __set_union, set_union, __set_intersection,
set_intersection, __set_difference, set_difference,
__set_symmetric_difference, set_symmetric_difference):  Mark constexpr.
* include/bits/stl_algobase.h (__memmove, __memcmp): New maybe constexpr
wrappers around __builtin_memmove and __builtin_memcmp
respectively;
(__niter_base, __niter_wrap, __copy_m, __copy_move_a, __copy_move_a2,
copy, move, __copy_move_b, __copy_move_backward_a,
__copy_move_backward_a2, copy_backward, move_backward, __fill_a, fill,
__fill_n_a, fill_n, equal, __lc_rai::__newlast1, __lc_rai::__cnd2,
__lexicographical_compare_impl, __lexicographical_compare,
__lexicographical_compare<true>::__lc, __lexicographical_compare_aux,
__lower_bound, lower_bound, equal, __equal4, lexicographical_compare,
__mismatch, mismatch, __is_heap_until, __is_heap, is_heap_until,
is_heap): Mark constexpr.
* include/bits/stl_heap.h (__is_heap_until, __is_heap, is_heap_until,
is_heap): Mark constexpr.
* include/bits/stl_iterator.h (__niter_base, __miter_base): Mark constexpr.
* include/std/array: Make comparison ops constexpr.
* include/std/utility: Make exchange constexpr.
* include/std/version (__cpp_lib_constexpr_algorithms): New macro.
* testsuite/23_containers/array/tuple_interface/get_neg.cc: Adjust.
* testsuite/23_containers/array/tuple_interface/
tuple_element_neg.cc: Adjust.
* testsuite/20_util/exchange/constexpr.cc: New.
* testsuite/23_containers/array/comparison_operators/constexpr.cc: New.
* testsuite/25_algorithms/constexpr_macro.cc: New.
* testsuite/25_algorithms/adjacent_find/constexpr.cc: New.
* testsuite/25_algorithms/all_of/constexpr.cc: New.
* testsuite/25_algorithms/any_of/constexpr.cc: New.
* testsuite/25_algorithms/binary_search/constexpr.cc: New.
* testsuite/25_algorithms/copy/constexpr.cc: New.
* testsuite/25_algorithms/copy_backward/constexpr.cc: New.
* testsuite/25_algorithms/copy_if/constexpr.cc: New.
* testsuite/25_algorithms/copy_n/constexpr.cc: New.
* testsuite/25_algorithms/count/constexpr.cc: New.
* testsuite/25_algorithms/count_if/constexpr.cc: New.
* testsuite/25_algorithms/equal/constexpr.cc: New.
* testsuite/25_algorithms/equal_range/constexpr.cc: New.
* testsuite/25_algorithms/fill/constexpr.cc: New.
* testsuite/25_algorithms/fill_n/constexpr.cc: New.
* testsuite/25_algorithms/find/constexpr.cc: New.
* testsuite/25_algorithms/find_end/constexpr.cc: New.
* testsuite/25_algorithms/find_first_of/constexpr.cc: New.
* testsuite/25_algorithms/find_if/constexpr.cc: New.
* testsuite/25_algorithms/find_if_not/constexpr.cc: New.
* testsuite/25_algorithms/for_each/constexpr.cc: New.
* testsuite/25_algorithms/generate/constexpr.cc: New.
* testsuite/25_algorithms/generate_n/constexpr.cc: New.
* testsuite/25_algorithms/is_heap/constexpr.cc: New.
* testsuite/25_algorithms/is_heap_until/constexpr.cc: New.
* testsuite/25_algorithms/is_partitioned/constexpr.cc: New.
* testsuite/25_algorithms/is_permutation/constexpr.cc: New.
* testsuite/25_algorithms/is_sorted/constexpr.cc: New.
* testsuite/25_algorithms/is_sorted_until/constexpr.cc: New.
* testsuite/25_algorithms/lexicographical_compare/constexpr.cc: New.
* testsuite/25_algorithms/lower_bound/constexpr.cc: New.
* testsuite/25_algorithms/merge/constexpr.cc: New.
* testsuite/25_algorithms/mismatch/constexpr.cc: New.
* testsuite/25_algorithms/none_of/constexpr.cc: New.
* testsuite/25_algorithms/partition_copy/constexpr.cc: New.
* testsuite/25_algorithms/partition_point/constexpr.cc: New.
* testsuite/25_algorithms/remove/constexpr.cc: New.
* testsuite/25_algorithms/remove_copy/constexpr.cc: New.
* testsuite/25_algorithms/remove_copy_if/constexpr.cc: New.
* testsuite/25_algorithms/remove_if/constexpr.cc: New.
* testsuite/25_algorithms/replace_copy/constexpr.cc: New.
* testsuite/25_algorithms/replace_copy_if/constexpr.cc: New.
* testsuite/25_algorithms/replace_if/constexpr.cc: New.
* testsuite/25_algorithms/reverse_copy/constexpr.cc: New.
* testsuite/25_algorithms/rotate_copy/constexpr.cc: New.
* testsuite/25_algorithms/search/constexpr.cc: New.
* testsuite/25_algorithms/search_n/constexpr.cc: New.
* testsuite/25_algorithms/set_difference/constexpr.cc: New.
* testsuite/25_algorithms/set_intersection/constexpr.cc: New.
* testsuite/25_algorithms/set_symmetric_difference/constexpr.cc: New.
* testsuite/25_algorithms/set_union/constexpr.cc: New.
* testsuite/25_algorithms/transform/constexpr.cc: New.
* testsuite/25_algorithms/unique/constexpr.cc: New.
* testsuite/25_algorithms/unique_copy/constexpr.cc: New.
* testsuite/25_algorithms/upper_bound/constexpr.cc: New.

From-SVN: r273975

Fix pr89330_0.C test

Fix pr89330_0.C test by adding missing effective target shared.
Committed as obvious.

   testsuite/
* gcc/testsuite/g++.dg/lto/pr89330_0.C: Add effective-target shared.

From-SVN: r273973

tree-ssa-pre.c (has_abnormal_preds): Remove global var.

2019-08-01  Richard Biener  <rguenther@suse.de>

* tree-ssa-pre.c (has_abnormal_preds): Remove global var.
(compute_antic): Localize it here.

From-SVN: r273971

Make-lang.in (lto.install-common): Remove unnecessary slash between $(DESTDIR) and $(bindir).

2019-08-01  Bernd Edlinger  <bernd.edlinger@hotmail.de>

        * Make-lang.in (lto.install-common): Remove unnecessary slash
        between $(DESTDIR) and $(bindir).

From-SVN: r273969

RISC-V: Fix testcase on rv64

gcc/testsuite/ChangeLog:

* gcc.target/riscv/attribute-10.c: Fix testcase on rv64.

From-SVN: r273968

Daily bump.

From-SVN: r273960

decl2.c (delete_sanity): Improve diagnostic locations, use cp_expr_loc_or_loc in four places.

/cp
2019-07-31  Paolo Carlini  <paolo.carlini@oracle.com>

* decl2.c (delete_sanity): Improve diagnostic locations, use
cp_expr_loc_or_loc in four places.

/testsuite
2019-07-31  Paolo Carlini  <paolo.carlini@oracle.com>

* g++.dg/diagnostic/delete1.C: New.

From-SVN: r273952

RISC-V: Raise error on unexpected ISA string at end.

2019-07-31  Maxim Blinov  <maxim.blinov@embecosm.com>
gcc/
* common/config/riscv/riscv-common.c: Check -march string ends
with null.
gcc/testsuite/
* gcc.target/riscv/attribute-10.c: New test.

From-SVN: r273951

Qualify call to prevent ADL

* include/std/memory (make_obj_using_allocator): Qualify call to
uses_allocator_construction_args.

From-SVN: r273945

PR c++/90538 - multiple expansions of capture packs

Previously, with init-capture the type of the closure field was a
DECLTYPE_TYPE of the initializer.  But since each time we tsubst a lambda we
get a different lambda, that meant that if the initializer is a lambda, we'd
end up with different closure types in the field and initializer after
substitution (PR 87322).  We dealt with this by remembering the lambda
instantiation within each pack expansion element, using
local_specialization_stack to separate the elements.  But that broke this
testcase, because it lost lambda capture proxies that also use
local_specializations.

So, this patch removes the local_specializations changes from that patch and
fixes 87322 differently, by giving init-capture fields 'auto' type and doing
deduction later.  There's a bit of a kludge to get the right number of
fields by pretending that 'auto...' uses the parameter packs from the
initializer, but it does the trick.

* cp-tree.h (DECLTYPE_FOR_INIT_CAPTURE): Remove.
* lambda.c (add_capture): Copy parameter packs from init.
(lambda_capture_field_type): Always use auto for init-capture.
* pt.c (uses_parameter_packs): Return tree.
(tsubst) [DECLTYPE_TYPE]: Remove init-capture handling.
(gen_elem_of_pack_expansion_instantiation): Don't push
local_specialization_stack.
(prepend_one_capture): New.
(tsubst_lambda_expr): Use it.  Don't touch local_specializations.
(do_auto_deduction): Avoid redundant error.

From-SVN: r273944

Fix copy_node of TEMPLATE_INFO.

build_clone uses copy_node to duplicate the TEMPLATE_INFO for a clone, but
this clears TREE_CHAIN, which was TI_ARGS in a TEMPLATE_INFO.

* cp-tree.h (struct tree_template_info): Use tree_base instead of
tree_common.  Add tmpl and args fields.
(TI_TEMPLATE, TI_ARGS): Adjust.

From-SVN: r273943

ipa-devirt: make qsort helpers static

* ipa-devirt.c (type_warning_cmp): Make static.
(decl_warning_cmp): Ditto.

From-SVN: r273942

re PR target/91050 (-mdejagnu-cpu=<cpu> does not affect the -m<cpu> assembler option)

PR target/91050
* config/rs6000/rs6000.opt (mdejagnu-cpu=): Delete option.
* config/rs6000/rs6000.c (rs6000_option_override_internal): Remove
use of deleted rs6000_dejagnu_cpu_index variable.
* config/rs6000/rs6000.h (DRIVER_SELF_SPECS): Define.
(SUBTARGET_DRIVER_SELF_SPECS): Likewise.
* config/darwin.h (DRIVER_SELF_SPECS): Rename from this ...
(SUBTARGET_DRIVER_SELF_SPECS): ...to this.
* config/i386/i386.h (DRIVER_SELF_SPECS): Define.
(SUBTARGET_DRIVER_SELF_SPECS): Likewise.

From-SVN: r273941

Implement "P0631R4 Math Constants" for C++20

The values of the constants are taken from Glibc where the equivalent
constant exists, or by rounding the actual constant to the same number
of digits as the Glibc constants have.

P0631R4 Math Constants
* include/Makefile.am: Add new header.
* include/Makefile.in: Regenerate.
* include/precompiled/stdc++.h: Include new header.
* include/std/numbers: New header.
* include/std/version (__cpp_lib_math_constants): Define.
* testsuite/26_numerics/numbers/1.cc: New test.
* testsuite/26_numerics/numbers/2.cc: New test.
* testsuite/26_numerics/numbers/3.cc: New test.
* testsuite/26_numerics/numbers/nonfloat_neg.cc: New test.

From-SVN: r273940

Add Doxygen comments to <bit> header

* include/std/bit: Add Doxygen comments.

From-SVN: r273938

PR libstdc++/91308 fix constraints on unique_ptr assignment

PR libstdc++/91308
* include/bits/unique_ptr.h (unique_ptr::__safe_conversion_up): Remove
constraints on deleter that should only apply to the constructor.
(unique_ptr<T[], D>::__safe_conversion_up): Likewise.
(unique_ptr<T[], D>::unique_ptr(unique_ptr<U, D>&&)): Restore
constraints on deleter here.
* testsuite/20_util/unique_ptr/assign/91308.cc: New test.

From-SVN: r273937

re PR tree-optimization/91280 (ICE in get_constraint_for_component_ref, at tree-ssa-structalias.c:3259 since r260354)

2019-07-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91280
* tree-ssa-structalias.c (get_constraint_for_component_ref):
Decompose MEM_REF manually for offset handling.

* g++.dg/torture/pr91280.C: New testcase.

From-SVN: r273936

re PR c/91192 (non-deterministic ICE on invalid)

PR c/91192
* c-parser.c (c_parser_sizeof_expression): Call set_c_expr_source_range
even if finish is UNKNOWN_LOCATION, just use start as finish in that
case.

From-SVN: r273935

re PR tree-optimization/91293 (Wrong code with -O3 -mavx2)

2019-07-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91293
* tree-vect-slp.c (vect_build_slp_tree_2): Do not swap operands
of reduction stmts.

* gcc.dg/vect/pr91293-1.c: New testcase.
* gcc.dg/vect/pr91293-2.c: Likewise.
* gcc.dg/vect/pr91293-3.c: Likewise.

From-SVN: r273934

config.gcc (hppa*-*-netbsd*): New target.

gcc/ChangeLog:
* config.gcc (hppa*-*-netbsd*): New target.
* config/pa/pa-netbsd.h: New file.
* config/pa/pa32-netbsd.h: New file.

libgcc/ChangeLog:
* config.host (hppa*-*-netbsd*): New case.
* config/pa/t-netbsd: New file.

Co-Authored-By: Matthew Green <mrg@eterna.com.au>
Co-Authored-By: Maya Rashish <coypu@sdf.org>
Co-Authored-By: Nick Hudson <nick@nthcliff.demon.co.uk>
From-SVN: r273933

re PR tree-optimization/91201 (SIMD not generated for horizontal sum of bytes in array)

PR tree-optimization/91201
* config/i386/mmx.md (reduc_plus_scal_v8qi): New expander.

* gcc.target/i386/sse2-pr91201-2.c: New test.

From-SVN: r273932

Remove amdgcn expcnt waits.

2019-07-31  Andrew Stubbs  <ams@codesourcery.com>

gcc/
* config/gcn/gcn-valu.md
(scatter<mode>_insn_1offset<exec_scatter>): Remove s_waitcnt.
(scatter<mode>_insn_1offset_ds<exec_scatter>): Likewise.
(scatter<mode>_insn_2offsets<exec_scatter>): Likewise.
* config/gcn/gcn.c (gcn_md_reorg): Add delayeduse and reads to
struct ilist. Add nops for delayeduse insns.
* config/gcn/gcn.md (delayeduse): New attribute.
(*movbi): Remove s_waitcnt from stores.
(*mov<mode>_insn): Likewise.
(*movti_insn): Likewise. Add delayeduse attribute.
(sync_compare_and_swap<mode>_insn): Add delayeduse attribute.
(atomic_store<mode>): Remove or adjust s_waitcnt.

From-SVN: r273931

vr-values.h (vr_values::swap_vr_value): New.

2019-07-31  Richard Biener  <rguenther@suse.de>

* vr-values.h (vr_values::swap_vr_value): New.
(vr_values::free_value_range): likewise.
* vr-values.c (vr_values::swap_vr_value): Implement.
* gimple-ssa-evrp-analyze.h (evrp_range_analyzer::pop_value_range):
Do not return a range or take a var.
(evrp_range_analyzer::stack): Change back to recording a non-const
value_range *.
* gimple-ssa-evrp-analyze.c
(evrp_range_analyzer::record_ranges_from_stmt): Free unused
value-range.
(evrp_range_analyzer::pop_to_marker): Adjust.
(evrp_range_analyzer::push_value_range): Use new swap_vr_value.
(evrp_range_analyzer::pop_value_range): Likewise.  Free the
no longer needed value-range.

From-SVN: r273930

Mark necessary 2nd and later args for delete op.

2019-07-31  Martin Liska  <mliska@suse.cz>

* tree-ssa-dce.c (propagate_necessity): Delete operator can
have size and (or) alignment as 2nd and later arguments.
Mark all of them as necessary.

From-SVN: r273929

re PR tree-optimization/91178 (Infinite recursion in split_constant_offset in slp after r260289)

2019-07-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91178
* tree-ssa-sccvn.c (vn_reference_maybe_forwprop_address):
Use tail-recursion.

* gcc.dg/torture/pr91178-2.c: New testcase.

From-SVN: r273928

re PR tree-optimization/91201 (SIMD not generated for horizontal sum of bytes in array)

PR tree-optimization/91201
* config/i386/sse.md (reduc_plus_scal_v16qi): New expander.
(REDUC_PLUS_MODE): Add V32QImode for TARGET_AVX and V64QImode for
TARGET_AVX512F.
(reduc_plus_scal_<mode>): Improve formatting by introducing
a temporary.

* gcc.target/i386/sse2-pr91201.c: New test.
* gcc.target/i386/avx2-pr91201.c: New test.
* gcc.target/i386/avx512bw-pr91201.c: New test.

From-SVN: r273927

[GCC, AArch64] Enable Transactional Memory Extension

This patch enables the new Transactional Memory Extension announced recently
as part of Arm's new architecture technologies.
We introduce a new optional extension "tme" to enable this. The following
instructions are part of the extension:
   * tstart <Xt>
   * ttest <Xt>
   * tcommit
   * tcancel #<imm>
We have also added ACLE intrinsics for the instructions.

*** gcc/ChangeLog ***

2019-07-31  Sudakshina Das  <sudi.das@arm.com>

* config/aarch64/aarch64-builtins.c (enum aarch64_builtins): Add
AARCH64_TME_BUILTIN_TSTART, AARCH64_TME_BUILTIN_TCOMMIT,
AARCH64_TME_BUILTIN_TTEST and AARCH64_TME_BUILTIN_TCANCEL.
(aarch64_init_tme_builtins): New.
(aarch64_init_builtins): Call aarch64_init_tme_builtins.
(aarch64_expand_builtin_tme): New.
(aarch64_expand_builtin): Handle TME builtins.
* config/aarch64/aarch64-c.c (aarch64_update_cpp_builtins): Define
__ARM_FEATURE_TME when enabled.
* config/aarch64/aarch64-option-extensions.def: Add "tme".
* config/aarch64/aarch64.h (AARCH64_FL_TME, AARCH64_ISA_TME): New.
(TARGET_TME): New.
* config/aarch64/aarch64.md (define_c_enum "unspec"): Add UNSPEC_TTEST.
(define_c_enum "unspecv"): Add UNSPECV_TSTART, UNSPECV_TCOMMIT and
UNSPECV_TCANCEL.
(tstart, ttest, tcommit, tcancel): New instructions.
* config/aarch64/arm_acle.h (__tstart, __tcommit): New.
(__tcancel, __ttest): New.
(_TMFAILURE_REASON, _TMFAILURE_RTRY, _TMFAILURE_CNCL): New macro.
(_TMFAILURE_MEM, _TMFAILURE_IMP, _TMFAILURE_ERR): Likewise.
(_TMFAILURE_SIZE, _TMFAILURE_NEST, _TMFAILURE_DBG): Likewise.
(_TMFAILURE_INT, _TMFAILURE_TRIVIAL): Likewise.
* config/arm/types.md: Add new tme type attr.
* doc/invoke.texi: Document "tme".

*** gcc/testsuite/ChangeLog ***

2019-07-31  Sudakshina Das  <sudi.das@arm.com>

* gcc.target/aarch64/acle/tme.c: New test.
* gcc.target/aarch64/pragma_cpp_predefs_2.c: New test.

From-SVN: r273926

[Arm][CMSE]Add warn_unused_return attribute to cmse functions

At present it is possible to call the CMSE functions for checking
addresses (such as cmse_check_address_range) and  forget to check/use
the return value. This patch makes the interfaces more robust against
programmer error by marking these functions with the warn_unused_result
attribute. With this set, any use of these functions that does not use
the result will produce a warning.

This produces a warning on default warn levels when the result of the
cmse functions is not used.

For the following function:
void foo()
{
     int *data;
     cmse_check_address_range((int*)data, 0, 0);
}
The following warning is emitted:
warning: ignoring return value of 'cmse_check_address_range' declared
with attribute 'warn_unused_result' [-Wunused-result]
     6 |  cmse_check_address_range((int*)data, 0, 0);
        |  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

gcc/ChangeLog:

2019-07-31  Joel Hutton  <Joel.Hutton@arm.com>

         * config/arm/arm_cmse.h (cmse_nonsecure_caller): Add
warn_unused_result attribute.
         (cmse_check_address_range): Add warn_unused_result attribute.

libgcc/ChangeLog:

2019-07-31  Joel Hutton  <Joel.Hutton@arm.com>

         * config/arm/cmse.c (cmse_check_address_range): Add
warn_unused_result attribute.

2019-07-31  Joel Hutton  <Joel.Hutton@arm.com>

         * gcc.target/arm/cmse/cmse-17.c: New test.

From-SVN: r273924

re PR tree-optimization/91257 (Compile-time and memory-hog hog)

2019-07-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91257
* tree-vrp.c (union_ranges): Unify equality and less tests
by using compare_values.  Re-order cheap tests first.

From-SVN: r273923

re PR middle-end/91301 (ICE in omp_add_variable on random access iterator distribute parallel for private (iterator))

PR middle-end/91301
* gimplify.c (gimplify_omp_for): If for class iterator on
distribute parallel for there is no data sharing clause
on inner_for_stmt, look for private clause on combined
parallel too and if found, move it to inner_for_stmt.

* testsuite/libgomp.c++/for-27.C: New test.

From-SVN: r273922

Make lra use per-alternative earlyclobber info

lra_insn_reg and lra_operand_data have both a bitmask of earlyclobber
alternatives and an overall boolean.  The danger is that we then test
the overall boolean when really we should be testing for a particular
alternative.  This patch gets rid of the boolean and tests the mask
against zero when we really do need to test "any alternative might
be earlyclobber".  (I think the only instance of that is the
LRA_UNKNOWN_ALT handling in lra-lives.c:reg_early_clobber_p.)

This is needed (and tested) by an upcoming SVE patch.

2019-07-31  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* lra-int.h (lra_operand_data): Remove early_clobber field.
(lra_insn_reg): Likewise.
* lra.c (debug_operand_data): Update accordingly.
(setup_operand_alternative): Likewise.
(new_insn_reg): Likewise.  Remove early_clobber parameter.
(collect_non_operand_hard_regs): Update call accordingly.
Don't assign to lra_insn_reg::early_clobber.
(add_regs_to_insn_regno_info): Remove early_clobber parameter
and update calls to new_insn_reg.
(lra_update_insn_regno_info): Update calls accordingly.
* lra-constraints.c (update_and_check_small_class_inputs): Take the
alternative number as a parameter and test whether the operand
is earlyclobbered in that particular alternative.
(process_alt_operands): Update call accordingly.  Use per-alternative
checks for earyclobber here too.
* lra-lives.c (reg_early_clobber_p): Check early_clobber_alts
against zero for IRA_UNKNOWN_ALT.

From-SVN: r273921

Daily bump.

From-SVN: r273920

PR testsuite/91258 - g++.dg/ubsan/vla-1.C and gcc.dg/strlenopt-70.c fail starting with r273783

gcc/testsuite/ChangeLog:
* g++.dg/ubsan/vla-1.C: Suppress a valid warning.

From-SVN: r273915

re PR fortran/91296 (ICE when passing complex number %re/%im as a procedure argument with -Waliasing.)

2019-07-30  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91296
* interface.c (compare_actual_expr): When checking for aliasing, add
a case to handle REF_INQUIRY (e.g., foo(x%re, x%im) do not alias).

2019-07-30  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/91296
* gfortran.dg/pr91296.f90: New test.

From-SVN: r273914

alpha.c (alpha_option_override): Quote a C type.

* config/alpha/alpha.c (alpha_option_override): Quote a C type.

From-SVN: r273912

Adjust literal pool offset in Thumb-2 movsi patterns

My previous change to the Thumb-2 movsi patterns caused a codesize regression
with -Os in large functions.  Fix this by using the literal pool offset of the
16-bit literal load so that the literal pool is dumped earlier, reducing the
number of 32-bit literal loads.

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

    gcc/
* config/arm/thumb2.md (thumb2_movsi_insn): Adjust literal offset.
* config/arm/vfp.md (thumb2_movsi_vfp): Likewise.

From-SVN: r273911

Use edge->indirect_unknown_callee in cgraph_edge::make_direct (PR ipa/89330).

2019-07-30  Martin Liska  <mliska@suse.cz>

PR ipa/89330
* cgraph.c (cgraph_edge::make_direct): Use
edge->indirect_unknown_callee as edge->resolve_speculation can
deallocate edge which is this pointer.

From-SVN: r273910

re PR tree-optimization/91257 (Compile-time and memory-hog hog)

2019-07-30  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91257
* bitmap.c (bitmap_ior_and_compl_into): Open-code.

From-SVN: r273909

Deduce automatically number of cores for -flto option.

2019-07-30  Martin Liska  <mliska@suse.cz>

* doc/invoke.texi: Document new behavior.
* lto-wrapper.c (cpuset_popcount): New function
is a copy of libgomp/config/linux/proc.c.
(init_num_threads): Likewise.
(run_gcc): Automatically detect core count for -flto.
(jobserver_active_p): New function.

From-SVN: r273908

re PR tree-optimization/91257 (Compile-time and memory-hog hog)

2019-07-30  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91257
* bitmap.h (bitmap_ior_into_and_free): Declare.
* bitmap.c (bitmap_list_unlink_element): Add defaulted param
whether to add the unliked element to the freelist.
(bitmap_list_insert_element_after): Add defaulted param for
an already allocated element.
(bitmap_ior_into_and_free): New function.
* tree-ssa-structalias.c (condense_visit): Reduce the
ponts-to and edge bitmaps of the SCC members in a
logarithmic fashion rather than all to one.

From-SVN: r273907

Mark 2nd argument of delete operator as needed (PR tree-optimization/91270).

2019-07-30  Martin Liska  <mliska@suse.cz>

PR tree-optimization/91270
* tree-ssa-dce.c (propagate_necessity): Mark 2nd argument
of delete operator as needed.
2019-07-30  Martin Liska  <mliska@suse.cz>

PR tree-optimization/91270
* g++.dg/torture/pr91270.C: New test.

From-SVN: r273906

Handle IFN_COND_MUL in tree-ssa-math-opts.c

This patch extends the FMA handling in tree-ssa-math-opts.c so
that it can cope with conditional multiplications as well as
unconditional multiplications.  The addition or subtraction must then
have the same condition as the multiplication (at least for now).

E.g. we can currently fold:

  (IFN_COND_ADD cond (mul x y) z fallback)
    -> (IFN_COND_FMA cond x y z fallback)

This patch also allows:

  (IFN_COND_ADD cond (IFN_COND_MUL cond x y <whatever>) z fallback)
    -> (IFN_COND_FMA cond x y z fallback)

2019-07-30  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-ssa-math-opts.c (convert_mult_to_fma): Add a mul_cond
parameter.  When nonnull, make sure that the addition or subtraction
has the same condition.
(math_opts_dom_walker::after_dom_children): Try convert_mult_to_fma
for CFN_COND_MUL too.

gcc/testsuite/
* gcc.dg/vect/vect-cond-arith-7.c: New test.

From-SVN: r273905

re PR tree-optimization/91291 (gcc.dg/20020312-2.c FAILs)

2019-07-30  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91291
* tree-ssa-sccvn.c (rpo_elim::eliminate_push_avail): Ignore
constant values.

From-SVN: r273903

re PR middle-end/91282 (gcc.dg/type-convert-var.c FAILs)

PR middle-end/91282
* gcc.dg/type-convert-var.c: Add -fexcess-precision=fast to
dg-additional-options.

From-SVN: r273899

re PR middle-end/91216 (OpenMP ICE starting with r265930)

PR middle-end/91216
* omp-low.c (global_nonaddressable_vars): New variable.
(use_pointer_for_field): For global decls, if they are non-addressable,
remember it in the global_nonaddressable_vars bitmap, if they are
addressable and in the global_nonaddressable_vars bitmap, ignore their
TREE_ADDRESSABLE bit.
(omp_copy_decl_2): Clear TREE_ADDRESSABLE also on private copies of
vars in global_nonaddressable_vars bitmap.
(execute_lower_omp): Free global_nonaddressable_vars bitmap.

* gcc.dg/gomp/pr91216.c: New test.

From-SVN: r273898

re PR target/91150 (wrong code with -O -mavx512vbmi due to wrong writemask)

PR target/91150
* config/i386/i386-expand.c (expand_vec_perm_blend): Change mask type
from unsigned to unsigned HOST_WIDE_INT.  For E_V64QImode cast
comparison to unsigned HOST_WIDE_INT before shifting it left.

* gcc.target/i386/avx512bw-pr91150.c: New test.

From-SVN: r273897

Daily bump.

From-SVN: r273896

i386.md (movstrict<mode>): Use register_operand predicate for operand 0.

* config/i386/i386.md (movstrict<mode>): Use register_operand
predicate for operand 0.  Add expander condition.  Assert that
operand 0 is a SUBREG RTX.
(*movstrict<mode>_1): Use register_operand predicate for operand 0.
Update operand constraints and insn condition.
(zero_extend<mode>si2_and): Do not call gen_movstrict<mode>.
(zero_extendqihi2_and): Do not call gen_movstrictqi.
(*setcc_qi_slp): Use register_operand predicate for operand 0.
Update operand 0 constraints.
(setcc_qi_slp splitters): Use register_operand predicate for operand 0.

From-SVN: r273891

MSP430: Disallow use of code/data regions in the small memory model

gcc/ChangeLog:

2019-07-29  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config/msp430/msp430.h (DRIVER_SELF_SPECS): Define and emit errors
when -m{code,data}-region are used without -mlarge.
* config/msp430/msp430.c (msp430_option_override): Error when a
non-default code or data region is used without -mlarge.
(msp430_section_attr): Emit a warning and do not add upper/lower/either
attributes when they are used without -mlarge.

gcc/testsuite/ChangeLog:

2019-07-29  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* gcc.target/msp430/pr78818-data-region.c: Add -mlarge to dg-options.
* gcc.target/msp430/region-misuse-code.c: New test.
* gcc.target/msp430/region-misuse-data.c: Likewise.
* gcc.target/msp430/region-misuse-code-data.c: Likewise.
* gcc.target/msp430/region-attribute-misuse.c: Likewise.

From-SVN: r273884

Allow both upper and lower case 'r' in register names

2019-07-29  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

PR target/70320
* config/msp430/msp430.h: Define ADDITIONAL_REGISTER_NAMES.

2019-07-29  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

PR target/70320
* gcc.target/msp430/asm-register-names-lower-case.c: New test.
* gcc.target/msp430/asm-register-names-upper-case.c: Likewise.

From-SVN: r273883

Add PR number

From-SVN: r273882

Fix inchash handling of wide_ints (PR91242)

inchash::hash::add_wide_int operated directly on the raw encoding
of the wide_int, including any redundant upper bits.  The problem
with that is that the upper bits are only defined for some wide-int
storage types (including wide_int itself).  wi::to_wide(tree) instead
returns a value that is extended according to the signedness of the
type (so that wi::to_widest can use the same encoding) while rtxes
have the awkward special case of BI, which can be zero-extended
rather than sign-extended.

In the PR, we computed a hash for a "normal" sign-extended wide_int
while the existing entries hashed wi::to_wide(tree).  This gives
different results for unsigned types that have the top bit set.

The patch fixes that by hashing the canonical sign-extended form even
if the raw encoding happens to be different.

2019-07-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* wide-int.h (generic_wide_int::sext_elt): New function.
* inchash.h (hash::add_wide_int): Use it instead of elt.

From-SVN: r273881

re PR fortran/90813 (gfortran.dg/proc_ptr_51.f90 fails (SIGSEGV) after 272084)

2019-07-29  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/90813
* dump-parse-tree.c (show_global_symbol): New function.
(gfc_dump_global_symbols): New function.
* gfortran.h (gfc_traverse_gsymbol): Add prototype.
(gfc_dump_global_symbols): Likewise.
* invoke.texi: Document -fdump-fortran-global.
* lang.opt: Add -fdump-fortran-global.
* parse.c (gfc_parse_file): Handle flag_dump_fortran_global.
* symbol.c (gfc_traverse_gsymbol): New function.
* trans-decl.c (sym_identifier): New function.
(mangled_identifier): New function, doing most of the work
of gfc_sym_mangled_identifier.
(gfc_sym_mangled_identifier): Use mangled_identifier.  Add mangled
identifier to global symbol table.
(get_proc_pointer_decl): Use backend decl from global identifier
if present.

From-SVN: r273880

[arm] Make ACLE builtins use arm_* namespace for expanders

The builtins from <arm_acle.h> use fairly general expander names such as
"crc", "mcr" etc.
These run the risk of being reserved by the midend in the future.
Let's namespace them to arm_* as is convention.

* config/arm/arm-builtins.c (acle_builtin_data): Expand VAR1 to
CODE_FOR_arm_##.
* config/arm/arm.md (<crc_variant>): Rename to...
(arm_<crc_variant>): ... This.
(<cdp>): Rename to...
(arm_<cdp>): ... This.
(<ldc>): Rename to...
(arm_<ldc>): ... This.
(<stc>): Rename to...
(arm_<stc>): ... This.
(<mcr>): Rename to...
(arm_<mcr>): ... This.
(<mrc>): Rename to...
(arm_<mrc>): ... This.
(<mcrr>): Rename to...
(arm_<mcrr>): ... This.
(<mrrc>): Rename to...
(arm_<mrrc>): ... This.

From-SVN: r273879

PR libstdc++/51333 Define recursive_init_error constructor non-inline

The recursive_init_error class is defined in a header, with an inline
constructor, but the definition of the vtable and destructor are not
exported from the shared library. With -fkeep-inline-functions the
constructor gets emitted in user code, and requires the (non-exported)
vtable. This fails to link.

As far as I can tell, the recursive_init_error class definition was
moved into <cxxabi.h> so it could be documented with Doxygen, not for
any technical reason. But now it's there (and documented), somebody
could be relying on it, by catching that type and possibly performing
derived-to-base conversions to the std::exception base class. So the
conservative fix is to leave the class definition in the header but make
the constructor non-inline. This still allows the type to be caught and
still defines its base class. User code can no longer construct objects
of that type, but that's not something we need to support.

PR libstdc++/51333
* libsupc++/cxxabi.h (__gnu_cxx::recursive_init_error): Do not define
constructor inline.
* libsupc++/guard_error.cc (__gnu_cxx::recursive_init_error): Define
constructor.
* testsuite/18_support/51333.cc: New test.

From-SVN: r273878

re PR tree-optimization/91257 (Compile-time and memory-hog hog)

2019-07-29  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91257
* tree-ssa-sccvn.h (struct vn_avail): New.
(struct vn_ssa_aux): Add avail member.
* tree-ssa-sccvn.c (class rpo_elim): Remove m_rpo_avail
member, add m_avail_freelist one.
(rpo_elim::~rpo_elim): Remove.
(rpo_elim::eliminate_avail): Adjust to new avail tracking
data structure.
(rpo_elim::eliminate_push_avail): Likewise.
(do_unwind): Likewise.
(do_rpo_vn): Likewise.

From-SVN: r273877

re PR tree-optimization/91257 (Compile-time and memory-hog hog)

2019-07-29  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91257
* tree-vrp.c (operand_less_p): Avoid dispatching to fold for
most cases, instead call compare_values which handles the
symbolic ranges we handle specially.
(compare_values_warnv): Do not call operand_less_p but open-code
the effective fold calls.  Avoid converting so much.

From-SVN: r273876

Fix ICE seen in tree-ssa-dce.c for new/delete pair.

2019-07-29  Martin Liska  <mliska@suse.cz>

* tree-ssa-dce.c (eliminate_unnecessary_stmts): Do not
remove LHS of operator new call.  It's handled latter.
2019-07-29  Martin Liska  <mliska@suse.cz>

* g++.dg/cpp1y/new1.C (test_unused): Add new case that causes
ICE.

From-SVN: r273875

re PR middle-end/91267 (SEGV in value_range_base::equal_p)

2019-07-29  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91267
* vr-values.c (vr_values::update_value_range): Add early return
for effectively VARYING lattice entry.

* gcc.dg/torture/pr91267.c: New testcase.

From-SVN: r273874

Prevent tree-ssa-dce.c from deleting stores at -Og

DCE tries to delete dead stores to local data and also tries to insert
debug binds for simple cases:

  /* If this is a store into a variable that is being optimized away,
     add a debug bind stmt if possible.  */
  if (MAY_HAVE_DEBUG_BIND_STMTS
      && gimple_assign_single_p (stmt)
      && is_gimple_val (gimple_assign_rhs1 (stmt)))
    {
      tree lhs = gimple_assign_lhs (stmt);
      if ((VAR_P (lhs) || TREE_CODE (lhs) == PARM_DECL)
  && !DECL_IGNORED_P (lhs)
  && is_gimple_reg_type (TREE_TYPE (lhs))
  && !is_global_var (lhs)
  && !DECL_HAS_VALUE_EXPR_P (lhs))
{
  tree rhs = gimple_assign_rhs1 (stmt);
  gdebug *note
    = gimple_build_debug_bind (lhs, unshare_expr (rhs), stmt);
  gsi_insert_after (i, note, GSI_SAME_STMT);
}
    }

But this doesn't help for things like "print *ptr" when ptr points
to the local variable (tests Og-dce-1.c and Og-dce-2.c).  It can
also introduce wrong debug info for earlier references (second test
in Og-dce-3.c) or make earlier references unavailable (first test
in Og-dce-3.c).

So for -Og I think it'd be better not to delete any stmts with
vdefs for now.  This also means that we can avoid the potentially
expensive vop walks (which already have a cut-off, but still).

The patch also fixes the Og failures in gcc.dg/guality/pr54970.c
(PR 86638).

2019-07-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR debug/86638
* tree-ssa-dce.c (keep_all_vdefs_p): New function.
(mark_stmt_if_obviously_necessary): Mark all stmts with vdefs as
necessary if keep_all_vdefs_p is true.
(mark_aliased_reaching_defs_necessary): Add a gcc_checking_assert
that keep_all_vdefs_p is false.
(mark_all_reaching_defs_necessary): Likewise.
(propagate_necessity): Skip the vuse scan if keep_all_vdefs_p is true.

gcc/testsuite/
* c-c++-common/guality/Og-dce-1.c: New test.
* c-c++-common/guality/Og-dce-2.c: Likewise.
* c-c++-common/guality/Og-dce-3.c: Likewise.

From-SVN: r273872

Don't run DSE at -Og

This patch stops gimple and rtl DSE from running by default at -Og.
The idea is both to improve compile time and to stop us from deleting
stores that we can't track in debug info.

We could rein this back in future for stores to local variables
with is_gimple_reg_type, but at the moment we don't have any
infrastructure for switching between binds to specific values
and binds to evolving memory locations.  Even then, location
tracking only works for direct references to the variables, and doesn't
for example help with printing dereferenced pointers (see the next patch
in the series for an example).

I'm also not sure that DSE is important enough for -Og to justify the
compile time cost --  especially in the case of RTL DSE, which is pretty
expensive.

2019-07-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* common.opt (Og): Change the initial value of flag_dse to 0.
* opts.c (default_options_table): Move OPT_ftree_dse from
OPT_LEVELS_1_PLUS to OPT_LEVELS_1_PLUS_NOT_DEBUG.  Also add
OPT_fdse to OPT_LEVELS_1_PLUS_NOT_DEBUG.  Put the OPT_ftree_pta
entry before the OPT_ftree_sra entry.
* doc/invoke.texi (Og): Add -fdse and -ftree-dse to the list
of flags disabled by Og.

gcc/testsuite/
* c-c++-common/guality/Og-global-dse-1.c: New test.

From-SVN: r273871

Prevent -Og from deleting stores to write-only variables

This patch prevents -Og from deleting stores to write-only variables,
so that the values are still available when debugging.  This seems
more convenient than forcing users to use __attribute__((used))
(probably conditionally, if it's not something they want in release
builds).

2019-07-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-cfg.c (execute_fixup_cfg): Don't delete stores to write-only
variables for -Og.

gcc/testsuite/
* c-c++-common/guality/Og-static-wo-1.c: New test.
* g++.dg/guality/guality.exp: Separate the c-c++-common tests into
"Og" and "general" tests.  Run the latter at -O0 and -Og only.
* gcc.dg/guality/guality.exp: Likewise.

From-SVN: r273870

Add dg test for matching function bodies

There isn't a 1:1 mapping from SVE intrinsics to SVE instructions,
but the intrinsics are still close enough to the instructions for
there to be a specific preferred sequence (or sometimes choice of
preferred sequences) for a given combination of operands.  Sometimes
these sequences will be one instruction, sometimes they'll be several.

I therefore wanted a convenient way of matching the exact assembly
implementation of a given function.  It's possible to do that using
single scan-assembler lines, but:

(a) they become hard to read for multiline matches
(b) the PASS/FAIL lines tend to be overly long
(c) it's useful to have a single place that skips over uninteresting
    lines, such as entry block labels and .cfi_* directives, without
    being overly broad

This patch therefore adds a new check-function-bodies dg-final test
that looks for specially-formatted comments.  As a demo, the patch
converts the SVE vec_init tests to use the new harness instead of
scan-assembler.

The regexps in parse_function_bodies are fairly general, but might
still need to be extended in future for targets like Darwin or AIX.

2019-07-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* doc/sourcebuild.texi (check-function-bodies): Document.

gcc/testsuite/
* lib/scanasm.exp (parse_function_bodies, check_function_body)
(check-function-bodies): New procedures.
* gcc.target/aarch64/sve/init_1.c: Use check-function-bodies
instead of scan-assembler.
* gcc.target/aarch64/sve/init_2.c: Likewise.
* gcc.target/aarch64/sve/init_3.c: Likewise.
* gcc.target/aarch64/sve/init_4.c: Likewise.
* gcc.target/aarch64/sve/init_5.c: Likewise.
* gcc.target/aarch64/sve/init_6.c: Likewise.
* gcc.target/aarch64/sve/init_7.c: Likewise.
* gcc.target/aarch64/sve/init_8.c: Likewise.
* gcc.target/aarch64/sve/init_9.c: Likewise.
* gcc.target/aarch64/sve/init_10.c: Likewise.
* gcc.target/aarch64/sve/init_11.c: Likewise.
* gcc.target/aarch64/sve/init_12.c: Likewise.

From-SVN: r273869

Generalise VEC_DUPLICATE folding for variable-length vectors

This patch uses the constant vector encoding scheme to handle
more cases of a VEC_DUPLICATE of another vector.  Duplicating
any fixed-length vector is fine, and duplicating a variable-length
vector is OK as long as that vector is also a duplicate of a
fixed-length sequence.

Other cases fell through to:

  if (VECTOR_MODE_P (mode) && GET_CODE (op) == CONST_VECTOR)

which was only expecting to deal with elementwise operations.

2019-07-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* simplify-rtx.c (simplify_const_unary_operation): Fold a
VEC_DUPLICATE of a fixed-length vector even if the result
is variable-length.  Likewise fold a duplicate of a
variable-length vector if the variable-length vector is
itself a duplicate of a fixed-length sequence.
(test_vector_ops_duplicate): Test more cases.

From-SVN: r273868

Implement more rtx vector folds on variable-length vectors

This patch extends the tree-level folding of variable-length vectors
so that it can also be used on rtxes.  The first step is to move
the tree_vector_builder new_unary/binary_operator routines to the
parent vector_builder class (which in turn means adding a new
template parameter).  The second step is to make simplify-rtx.c
use a direct rtx analogue of the VECTOR_CST handling in fold-const.c.

2019-07-29  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* vector-builder.h (vector_builder): Add a shape template parameter.
(vector_builder::new_unary_operation): New function, generalizing
the old tree_vector_builder function.
(vector_builder::new_binary_operation): Likewise.
(vector_builder::binary_encoded_nelts): Likewise.
* int-vector-builder.h (int_vector_builder): Update template
parameters to vector_builder.
(int_vector_builder::shape_nelts): New function.
* rtx-vector-builder.h (rtx_vector_builder): Update template
parameters to vector_builder.
(rtx_vector_builder::shape_nelts): New function.
(rtx_vector_builder::nelts_of): Likewise.
(rtx_vector_builder::npatterns_of): Likewise.
(rtx_vector_builder::nelts_per_pattern_of): Likewise.
* tree-vector-builder.h (tree_vector_builder): Update template
parameters to vector_builder.
(tree_vector_builder::shape_nelts): New function.
(tree_vector_builder::nelts_of): Likewise.
(tree_vector_builder::npatterns_of): Likewise.
(tree_vector_builder::nelts_per_pattern_of): Likewise.
* tree-vector-builder.c (tree_vector_builder::new_unary_operation)
(tree_vector_builder::new_binary_operation): Delete.
(tree_vector_builder::binary_encoded_nelts): Likewise.
* simplify-rtx.c: Include rtx-vector-builder.h.
(distributes_over_addition_p): New function.
(simplify_const_unary_operation)
(simplify_const_binary_operation): Generalize handling of vector
constants to include variable-length vectors.
(test_vector_ops_series): Add more tests.

From-SVN: r273867

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

PR lto/91222
* ipa-devirt.c (warn_types_mismatch): Compare indentifiers
than INDENTIFIER_POINTER.

From-SVN: r273866

Daily bump.

From-SVN: r273864

Release cgraph_{node,edge} via ggc_free (PR ipa/89330).

2019-07-28  Martin Liska  <mliska@suse.cz>

PR ipa/89330
* cgraph.c (symbol_table::create_edge): Always allocate
a cgraph_edge.
(symbol_table::free_edge): Store summary_id to
edge_released_summary_ids if != -1;
* cgraph.h (NEXT_FREE_NODE): Remove.
(SET_NEXT_FREE_NODE): Likewise.
(NEXT_FREE_EDGE): Likewise.
(symbol_table::release_symbol): Store summary_id to
cgraph_released_summary_ids if != -1;
(symbol_table::allocate_cgraph_symbol): Always allocate
a cgraph_node.

From-SVN: r273857

Fix g++.dg/lto/pr89330 on Solaris

* g++.dg/lto/pr89330_0.C (dg-lto-options): Add -fPIC.
Require fpic support.

From-SVN: r273856

[RS6000] rs6000_output_mi_thunk tidy

Since svn r123193 (git 327202e27a) the comment about scratch regs was
no longer correct, and I think gen_sibcall could have been used even
then.

* gcc/config/rs6000/rs6000-call.c (rs6000_output_mi_thunk): Use
gen_sibcall.

From-SVN: r273855

[RS6000] PR91135, __linux__ not defined with -mcall-aixdesc on 9.x and ppc64

This patch makes the obvious fix for PR91135, and deletes extraneous
copies of GNU_USER_TARGET_D_OS_VERSIONS that appear in rs6000/linux.h
and rs6000/linux64.h.  Since all configurations using either of these
files also include linux.h there is no need to duplicate the macro.

PR target/91135
* config/rs6000/linux.h (GNU_USER_TARGET_D_OS_VERSIONS): Don't
define.
* config/rs6000/linux64.h (TARGET_OS_CPP_BUILTINS): Invoke
GNU_USER_TARGET_OS_CPP_BUILTINS for aixdesc abi.
(GNU_USER_TARGET_D_OS_VERSIONS): Don't define.

From-SVN: r273854

[RS6000] Make assembler command line cpu match default for gcc

When gcc is configured using --with-cpu=<cpu>, the specified cpu
effectively becomes a default -mcpu=<cpu> passed to gcc.  This then
affects the cpu passed to gas via ASM_CPU_SPEC.  If gcc is not
configured using --with-cpu then the cpu passed to gas is that given
by ASM_DEFAULT_SPEC, which currently does not match the default flags
selected in default64.h.  This patch makes ASM_DEFAULT_SPEC agree with
TARGET_DEFAULT flags.

rs6000/default64.h appears in three places in config.gcc, the first
one immediately followed by rs6000/freebsd64.h in $tm_file, and the
other two immediately followed by rs6000/linux64.h.  To be able to
define ASM_DEFAULT_SPEC in rs6000/default64.h we don't want to
redefine in the other two files.  rs6000/freebsd64.h is easy since
that file is always preceded by rs6000/default64.h, but
rs6000/linux64.h can appear without rs6000/default64.h (a
powerpc*-linux config where the default is -m32).  In that case we
will have TARGET_DEFAULT flags of 0 (from rs6000/sysv4.h) and want to
use -mppc without -m64 and -mppc64 with -m64.  This can be done by
using the rs6000/rtems.h ASM_DEFAULT_SPEC in rs6000/sysv4.h, a change
that won't affect sysv4 configurations where -m64 is invalid.

The patch also introduces ASM_DEFAULT_EXTRA for the altivec variant
targets so as to enable -maltivec by default.

PR target/91050
* config/rs6000/sysv4.h (ASM_DEFAULT_SPEC): Modify if -m64.
* config/rs6000/default64.h (ASM_DEFAULT_SPEC): Define.
* config/rs6000/freebsd64.h (ASM_DEFAULT_SPEC): Don't define.
* config/rs6000/linux64.h (ASM_DEFAULT_SPEC): Likewise.
* config/rs6000/rtems.h (ASM_DEFAULT_SPEC): Likewise.
* config/rs6000/rs6000.h (ASM_DEFAULT_EXTRA): Define and use
in asm_default spec.
* config/rs6000/eabialtivec.h (ASM_DEFAULT_EXTRA): Redefine.
* config/rs6000/linuxaltivec.h (ASM_DEFAULT_EXTRA): Redefine.

From-SVN: r273853

documentation_hacking.xml: Fix broken reference to the Doxygen manual.

* doc/xml/manual/documentation_hacking.xml: Fix broken reference
to the Doxygen manual. Avoid a "here" link on the way.
Fix another broken link to Doxygen docblocks.

From-SVN: r273852

* doc/include/gpl_v3.texi (Copying): Use https for gnu.org.

From-SVN: r273851

Daily bump.

From-SVN: r273850

[Darwin, PPC, testsuite] Fix fail for bmi2-bzhi64-1a.c

This test is failing with older cpus because the included header needs both
altivec and vsx to be enabled to succeed in compiling.  Without this (if these
are not defaults for the cpu) there are errors like:

In file included from ... x86intrin.h:41,
                 from ... bmi2-bzhi64-1a.c:6:
 ... xmmintrin.h: In function '_mm_loadu_ps':
 ... xmmintrin.h:122:11:
error: incompatible types when returning type 'int' but '__m128' {aka '__vector(4) float'} was expected
<snip>
 ... xmmintrin.h: In function '_mm_cvtps_pi32':
 ... xmmintrin.h:996:3:
error: use of 'long long' in AltiVec types is invalid without '-mvsx'
<snip>

Fixed by adding -maltivec -mvsx to the options.

gcc/testsuite/

2019-07-27  Iain Sandoe  <iain@sandoe.co.uk>

* gcc.target/powerpc/bmi2-bzhi64-1a.c: Add options to enable altivec
and vsx.

From-SVN: r273844

Daily bump.

From-SVN: r273843

[Darwin, testsuite] Handle Darwin's size command.

Darwin's "size" command has a different header line, reflecting the Mach-O
section naming conventions.  This causes tests using the command to fail
because scanasm.exp expects and checks specific layout of the header line.

gcc/testsuite/

2019-07-26  Iain Sandoe  <iain@sandoe.co.uk>

* lib/scanasm.exp (object-size): Handle Darwin's size command.

From-SVN: r273839

Define __cpp_lib_endian feature test macro

This macro was added as part of moving std::endian from <type_traits> to
<bit>.

* include/std/bit (__cpp_lib_endian): Define.
* include/std/version (__cpp_lib_endian): Define.
* testsuite/26_numerics/endian/2.cc: New.
* testsuite/26_numerics/endian/3.cc: New.
* testsuite/26_numerics/endian/4.cc: New.

From-SVN: r273828

AArch64: Make processing less fragile in config.gcc

Due to config.gcc all the options need to be on one line because of the grep
lines which would select only the first line of the option.

This causes it not to select the right bits on options that are spread over
multiple lines when the --with-arch configure option is used.  The issue happens
silently and you just get a compiler with an incorrect set of default flags.

The current rules are quite rigid:

   1) No space between the AARCH64_OPT_EXTENSION and the opening (.
   2) No space between the opening ( and the extension name.
   3) No space after the extension name before the ,.
   4) Spaces are only allowed after a , and around |.

This patch makes this a lot less fragile by using the C pre-processor to flatten
the list and then provides much more flexible regex using group matching to
process the options instead of string replacement.  This removes all the
restrictions above and makes the code a bit more readable.

gcc/ChangeLog:

PR target/89517
* config.gcc: Relax parsing of AARCH64_OPT_EXTENSION.
* config/aarch64/aarch64-option-extensions.def: Add new comments
and restore easier to read options.

From-SVN: r273827

Add rules to strip away unneeded type casts in expressions

This patch moves part of the type conversion code from convert.c to match.pd
because match.pd is able to apply these transformations in the presence of
intermediate temporary variables.

Concretely it makes both these cases behave the same

  float e = (float)a * (float)b;
  *c = (_Float16)e;

and

  *c = (_Float16)((float)a * (float)b);

gcc/ChangeLog:

* convert.c (convert_to_real_1): Move part of conversion code...
* match.pd: ...To here.

gcc/testsuite/ChangeLog:

* gcc.dg/type-convert-var.c: New test.

From-SVN: r273826

[PR 89330] Remove non-useful speculations from new_edges

2019-07-26  Martin Jambor  <mjambor@suse.cz>

PR ipa/89330
* ipa-inline-transform.c (check_speculations_1): New function.
(push_all_edges_in_set_to_vec): Likewise.
(check_speculations): Use check_speculations_1, new parameter
new_edges.
(inline_call): Pass new_edges to check_speculations.
* ipa-inline.c (add_new_edges_to_heap): Assert edge_callee is not
NULL.
(speculation_useful_p): Early return true if edge is inlined, remove
later checks for inline_failed.

testsuite/
* g++.dg/lto/pr89330_[01].C: New test.
* g++.dg/tree-prof/devirt.C: Added -fno-profile-values to dg-options.

From-SVN: r273825

2019-07-26  François Dumont  <fdumont@gcc.gnu.org>

* testsuite/util/testsuite_iterators.h
(bidirectional_iterator_wrapper): Fix type comment.
(random_access_iterator_wrapper): Likewise.

From-SVN: r273824

Daily bump.

From-SVN: r273823

Relocate std::endian from <type_traits> to <bit>

This change to an early C++2a feature was just approved (P1612R1).

* include/std/bit (endian): Move definition here as per P1612R1.
* include/std/type_traits (endian): Remove definition from here.
* testsuite/20_util/endian/1.cc: Rename to ...
* testsuite/26_numerics/endian/1.cc: ... here. Adjust header.

From-SVN: r273816

Wstringop-overflow-14.c: Disable for stricly aligned targets.

gcc/testsuite/ChangeLog:
* gcc.dg/Wstringop-overflow-14.c: Disable for stricly aligned targets.

From-SVN: r273814

re PR tree-optimization/91183 (strlen of a strcpy result with a conditional source not folded)

PR tree-optimization/91183
PR tree-optimization/86688

gcc/testsuite/ChangeLog:
* gcc.dg/strlenopt-70.c: Fix bugs.
* gcc.dg/strlenopt-71.c: Same.
* gcc.dg/strlenopt-72.c: Same.

From-SVN: r273812

re PR rtl-optimization/91223 (ICE: in curr_insn_transform, at lra-constraints.c:4459)

2019-07-25  Vladimir Makarov  <vmakarov@redhat.com>

PR rtl-optimization/91223
* lra-constraints.c (process_alt_operands): Fail for unsuccessful
matching with INOUT operand.

2019-07-25  Vladimir Makarov  <vmakarov@redhat.com>

PR rtl-optimization/91223
* gcc.target/i386/pr91223.c: New test.

From-SVN: r273810

[Darwin, testsuite] Address PR91087 - XFAIL parts of pr16855.C.

The testcase is failing to instrument part of the source because of a bug
in the ordering of static DTORs.  It seems unlikely that this is generically
fixable in the toolchain (and given that it's likely to be a dynamic loader
change would not be expected to be applied retrospectively to OS versions
that are out of support).  To avoid the testsuite noise, xfail the count lines
that don't match (we can adjust the xfails as/when the upstream bug is fixed).

dejagnu xfails do not seem to work when embedded in a line like:
  ~Test (void) { .... /* count(1) { xfail ... } */ }
the closing brace seems to confuse the parser.  The solution is to exapnd the
text onto three lines.

2019-07-25  Iain Sandoe  <iain@sandoe.co.uk>

PR gcov-profile/91087
* g++.dg/gcov/pr16855.C: Xfail the count lines for the DTORs and the
"final" line for the failure summaries.  Adjust source layout so that
dejagnu xfail expressions work.

From-SVN: r273809

re PR fortran/65819 (overzealous checking in gfc_check_dependency for identical=true)

2019-07-25  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/65819
* dependency.h (gfc_dep_resovler): Add optional argument identical.
* dependency.c (gfc_check_dependency): Do not alway return 1 if
the symbol is the same. Pass on identical to gfc_dep_resolver.
(gfc_check_element_vs_element): Whitespace fix.
(gfc_dep_resolver): Adjust comment for function.  If identical is
true, return 1 if any overlap has been found.

2019-07-25  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/65819
* gfortran.dg/dependency_54.f90: New test.

From-SVN: r273807

stmt.c (expand_case): Try to narrow the index type if it's larger than a word.

* stmt.c (expand_case): Try to narrow the index type if it's larger
than a word.  Tidy up.

From-SVN: r273805

cif-code.def (NEVER_CALL): New code.

* cif-code.def (NEVER_CALL): New code.
* ipa-inline.c (want_inline_small_function_p): Fix formatting issues.
Set the failure to CIF_NEVER_CALL if the IPA count is zero.

From-SVN: r273804

Fix low reg issue in Thumb-2 movsi patterns

The Thumb-2 movsi patterns try to prefer low registers for loads and stores.
However this is done incorrectly by using 2 separate variants with 'l' and 'h'
register classes.  The register allocator will only use low registers, and
as a result we end up with significantly more spills and moves to high
registers.  Fix this by merging the alternatives and use 'l*r' to indicate
preference for low registers.  This saves ~400 instructions from the pr77308
testcase.

    gcc/
* config/arm/thumb2.md (thumb2_movsi_insn): Fix load/store low reg.
* config/arm/vfp.md (thumb2_movsi_vfp): Likewise.

From-SVN: r273802

* ipa-devirt.c (add_type_duplicate): Fix return value.

From-SVN: r273800

tree-vrp.c (extract_range_from_multiplicative_op): Add type parameter and use it instead of guessing expression type from the...

2019-07-25  Richard Biener  <rguenther@suse.de>

* tree-vrp.c (extract_range_from_multiplicative_op): Add
type parameter and use it instead of guessing expression
type from the first operand.
(extract_range_from_binary_expr): Pass expr_type down.

From-SVN: r273797

[arm][committed] Clean up code iterator usage in satsi* patterns

GCC 10 now supports having RTL codes being code attributes (thanks
Richard) allowing us to map smax to smin and vice versa.
This means we can clean up their use in the saturation patterns that do
the cross product of [smin, smax] and use the pattern
predicate to cancel out the nonsense ones.

* config/arm/arm.md (SATrev): Change to code attribute.
(*satsi_<SAT:code>): Adjust for the above.
(*satsi_<SAT:code>_shift): Likewise.

From-SVN: r273796