vect: Fix epilogue loop handling of partial vectors

This patch fixes the fallout that Kewen reported on Power after
the recent change to avoid unnecessary use of partial vectors.
As Kewen said, the problem is that vect_analyze_loop_2 doesn't
know how many epilogue iterations there will be, and so it
cannot make a final decision about whether the number of
iterations forces an epilogue loop to use partial vectors.

This is similar to the current situation for peeling: we don't know
during initial analysis whether an epilogue loop will itself require
peeling.  Instead we decide that during vect_do_peeling, where the
final number of epilogue loop iterations is known.

The patch takes a similar approach for the decision about whether
to use partial vectors.  As the comments in the patch say, the
idea is that vect_analyze_loop_2 should make peeling and partial-
vector decisions based on the assumption that the loop_vinfo will
be used as the main loop, while vect_do_peeling should make them
in the knowledge that the loop_vinfo will be used as an epilogue loop.

This allows the same analysis to be used for both cases, which we
rely on for implementing VECT_COMPARE_COSTS; see the big comment
in vect_analyze_loop for details.

I hope the patch makes the (mostly preexisting) structure a bit
more obvious.  It isn't what anyone would design from scratch,
but that's the nature of working with a mature vector framework.

Arranging things this way means that vect_verify_full_masking
and vect_verify_loop_lens now become part of the “can” rather
than “will” test for partial vectors.

Also, while splitting out the logic that handles epilogues with
constant iterations, I added a check to make sure that we don't
try to use partial vectors to vectorise a single-scalar loop.
This required some changes to the Power tests.

gcc/
* tree-vectorizer.h (determine_peel_for_niter): Delete in favor of...
(vect_determine_partial_vectors_and_peeling): ...this new function.
* tree-vect-loop-manip.c (vect_update_epilogue_niters): New function.
Reject using vector epilogue loops for single iterations.  Install
the constant number of epilogue loop iterations in the associated
loop_vinfo.  Rely on vect_determine_partial_vectors_and_peeling
to do the main part of the test.
(vect_do_peeling): Use vect_update_epilogue_niters to handle
epilogue loops with a known number of iterations.  Skip recomputing
the number of iterations later in that case.  Otherwise, use
vect_determine_partial_vectors_and_peeling to decide whether the
epilogue loop needs to use partial vectors or peeling.
* tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Set the
default can_use_partial_vectors_p to false if partial-vector-usage=0.
(determine_peel_for_niter): Remove in favor of...
(vect_determine_partial_vectors_and_peeling): ...this new function,
split out from...
(vect_analyze_loop_2): ...here.  Reflect the vect_verify_full_masking
and vect_verify_loop_lens results in CAN_USE_PARTIAL_VECTORS_P
rather than USING_PARTIAL_VECTORS_P.

gcc/testsuite/
* gcc.target/powerpc/p9-vec-length-epil-1.c: Do not expect the
single-iteration epilogues of the 64-bit loops to be vectorized.
* gcc.target/powerpc/p9-vec-length-epil-7.c: Likewise.
* gcc.target/powerpc/p9-vec-length-epil-8.c: Likewise.

AArch64: Implement missing vrndns_f32 intrinsic

This patch implements the missing vrndns_f32 intrinsic. This operates on a scalar float32_t value.
It can be mapped down to a __builtin_aarch64_frintnsf builtin.

This patch does that.

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

gcc/
PR target/71233
* config/aarch64/aarch64-simd-builtins.def (frintn): Use BUILTIN_VHSDF_HSDF
for modes.  Remove explicit hf instantiation.
* config/aarch64/arm_neon.h (vrndns_f32): Define.

gcc/testsuite/
PR target/71233
* gcc.target/aarch64/simd/vrndns_f32_1.c: New test.

tree-optimization/97173 - extend assert in vectorizable_live_operation

The condition we're expecting to eventually run into isn't fully
captured by checking for CTORs, instead we can also run into the
CTOR element conversion.

2020-09-23  Richard Biener  <rguenther@suse.de>

PR tree-optimization/97173
* tree-vect-loop.c (vectorizable_live_operation): Extend
assert to also conver element conversions.

* gcc.dg/vect/pr97173.c: New testcase.

AArch64: Implement missing _p64 intrinsics for vector permutes

This patch implements some missing vector permute intrinsics operating on poly64x2_t types.
They are implemented identically to their uint64x2_t brethren.

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

gcc/
PR target/71233
* config/aarch64/arm_neon.h (vtrn1q_p64, vtrn2q_p64, vuzp1q_p64,
vuzp2q_p64, vzip1q_p64, vzip2q_p64): Define.

gcc/testsuite/
PR target/71233
* gcc.target/aarch64/simd/trn_zip_p64_1.c: New test.

AArch64: Implement vldrq_p128 intrinsic

This patch implements the missing vldrq_p128 intrinsic that just loads from the appropriate pointer.

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

gcc/
PR target/71233
* config/aarch64/arm_neon.h (vldrq_p128): Define.

gcc/testsuite/
PR target/71233
* gcc.target/aarch64/simd/vldrq_p128_1.c: New test.

AArch64: Implement vstrq_p128 intrinsic

This patch implements the missing vstrq_p128 intrinsic.
It just performs a store of the poly128_t argument to a memory location.

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

gcc/
PR target/71233
* config/aarch64/arm_neon.h (vstrq_p128): Define.

gcc/testsuite/
PR target/71233
* gcc.target/aarch64/simd/vstrq_p128_1.c: New test.

gcc/analyzer: Silence -Wpragma warns with GCC < 10

gcc/analyzer/ChangeLog:

* analyzer-logging.cc: Guard '#pragma ... ignored "-Wformat-diag"'
by '#if __GNUC__ >= 10'
* analyzer.h: Likewise.
* call-string.cc: Likewise.

tree-optimization/97151 - improve PTA for C++ operator delete

C++ operator delete, when DECL_IS_REPLACEABLE_OPERATOR_DELETE_P,
does not cause the deleted object to be escaped.  It also has no
other interesting side-effects for PTA so skip it like we do
for BUILT_IN_FREE.

2020-09-23  Richard Biener  <rguenther@suse.de>

PR tree-optimization/97151
* tree-ssa-structalias.c (find_func_aliases_for_call):
DECL_IS_REPLACEABLE_OPERATOR_DELETE_P has no effect on
arguments.

* g++.dg/cpp1y/new1.C: Adjust for two more handled transforms.

middle-end/97162 - fix ICE when building gamess

This appropriately guards the check for a hard register in
compare_base_decls which otherwise ICEs when passed a CONST_DECL.

2020-09-23  Richard Biener  <rguenther@suse.de>

PR middle-end/97162
* alias.c (compare_base_decls): Use DECL_HARD_REGISTER
and guard with VAR_P.

gcov: fix streaming corruption

gcc/ChangeLog:

PR gcov-profile/97069
* profile.c (branch_prob): Line number must be at least 1.

gcc/testsuite/ChangeLog:

PR gcov-profile/97069
* g++.dg/gcov/pr97069.C: New test.

[nvptx] Handle move from DF subreg to DF reg in nvptx_output_mov_insn

When compiling test-case gcc.dg/atomic/c11-atomic-exec-1.c, we run into
these ptxas errors:
...
line 100; error: Rounding modifier required for instruction 'cvt'
line 105; error: Rounding modifier required for instruction 'cvt'
...

The problem is that this move:
...
//(insn 13 11 14 2
//      (set (reg:DF 28 [ _9 ])
//           (subreg:DF (reg:TI 22 [ _1 ]) 0)) 9 {*movdf_insn}
//       (nil))
                cvt.f64.u64     %r28, %r22$0;
...
is emitted as cvt.f64.u64, while it should be a mov.b64 instead.

Fix this by handling this case in nvptx_output_mov_insn.

Tested on nvptx.

gcc/ChangeLog:

PR target/97158
* config/nvptx/nvptx.c (nvptx_output_mov_insn): Handle move from
DF subreg to DF reg.

[testsuite] Add missing require-effective-target alloca

Add missing require-effect-target alloca directives.

Tested on nvptx.

gcc/testsuite/ChangeLog:

* gcc.dg/Warray-bounds-63.c: Add require-effective-target alloca.
* gcc.dg/Warray-bounds-66.c: Same.
* gcc.dg/atomic/stdatomic-vm.c: Same.

syscall: fix TestForeground for AIX

Syscall function can't be used on AIX. Therefore, Ioctl in
TestForeground must call raw_ioctl.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/175080

syscall: remove ptrace syscall on ppc64

ptrace is available only for 32 bits programs.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/252558

Add $(ZLIBINC) to CFLAGS-analyzer/engine.o

gcc/ChangeLog:
* Makefile.in: Add $(ZLIBINC) to CFLAGS-analyzer/engine.o.

analyzer: use switch in exploded_node::on_stmt

This patch replaces a sequence of dyn_cast to different gimple stmt
types in exploded_node::on_stmt with a switch on the gimple_code.  This
makes clearer which kinds of stmt are currently treated as no-ops, as a
precursor to handling them properly.

No functional change intended.

gcc/analyzer/ChangeLog:
* engine.cc (exploded_node::on_stmt): Replace sequence of dyn_cast
with switch.

runtime, net: fix build errors on AIX

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/235158

libbacktrace: handle pc == low correctly

* dwarf.c (report_inlined_functions): Handle PC == -1 and PC ==
p->low.
(dwarf_lookup_pc): Likewise.

Daily bump.

go.test: update issue4458.go for recent change

Ignore clobbers in modref

* ipa-modref.c (analyze_stmt): Ignore gimple clobber.

c++: Return only in-scope tparms in keep_template_parm [PR95310]

In the testcase below, the dependent specializations iter_reference_t<F>
and iter_reference_t<Out> share the same tree due to specialization
caching.  So when find_template_parameters walks through the
requires-expression (as part of normalization), it sees and includes the
out-of-scope template parameter F in the list of template parameters
it found within the requires-expression (along with Out and N).

From a correctness perspective this is harmless since the parameter mapping
routines only care about the level and index of each parameter, so F is
no different from Out in that sense.  And it's also harmless that two
parameters in the parameter mapping have the same level and index.

But having both Out and F in the parameter mapping means extra work for
hash_atomic_constrant, tsubst_parameter_mapping and get_mapped_args; and
it also means we print this irrelevant template parameter in the
testcase's diagnostics (via pp_cxx_parameter_mapping):

  in requirements with ‘Out o’ [with N = (const int&)&a; F = const int*; Out = const int*]

This patch makes keep_template_parm return only in-scope template
parameters by looking into ctx_parms for the corresponding in-scope
one, through a new helper function corresponding_template_parameter.

(That we sometimes print irrelevant template parameters in diagnostics
is also the subject of PR99 and PR66968, so the above diagnostic issue
could likely be fixed in a more general way, but this targeted fix to
keep_template_parm is perhaps worthwhile on its own.)

gcc/cp/ChangeLog:

PR c++/95310
* pt.c (corresponding_template_parameter): Define.
(keep_template_parm): Use it to adjust the given template
parameter to the corresponding in-scope one from ctx_parms.

gcc/testsuite/ChangeLog:

PR c++/95310
* g++.dg/concepts/diagnostic15.C: New test.

c++: Add test for PR96652

Fixed by r11-3361.

gcc/testsuite/ChangeLog:

PR c++/96652
* g++.dg/cpp0x/decltype-96652.C: New test.

Fix ipa-modref selftest and destructor

* ipa-modref-tree.c: Add namespace selftest.
(modref_tree_c_tests): Rename to ...
(ipa_modref_tree_c_tests): ... this.
* ipa-modref.c (pass_modref): Remove destructor.
(ipa_modref_c_finalize): New function.
* ipa-modref.h (ipa_modref_c_finalize): Declare.
* selftest-run-tests.c (selftest::run_tests): Call
ipa_modref_c_finalize.
* selftest.h (ipa_modref_tree_c_tests): Declare.
* toplev.c: Include ipa-modref-tree.h and ipa-modref.h
(toplev::finalize): Call ipa_modref_c_finalize.

c++: Remove a broken error-recovery path

The remaining use of xref_tag_from_type was also suspicious.  It turns
out to be an error path.  At parse time we diagnose that a class
definition cannot appear, but we swallow the definition.  This code
was attempting to push it into the global scope (or find a conflict).
This seems needless, just return error_mark_node.  This was the
simpler fix than going through the parser and figuring out how to get
it to put in error_mark_node at the right point.

gcc/cp/
* cp-tree.h (xref_tag_from_type): Don't declare.
* decl.c (xref_tag_from_type): Delete.
* pt.c (lookup_template_class_1): Erroneously located class
definitions just give error_mark, don't try and inject it into the
namespace.

c++: Ignore __sanitizer_ptr_{sub,cmp} builtin calls during constant expression evaluation [PR97145]

These two builtin calls are added already during parsing before pointer
subtractions or comparisons, normally they perform runtime verification
of whether the pointers point to the same object or different objects,
but during constant expressione valuation we don't really need those
builtins for anything.

2020-09-22  Jakub Jelinek  <jakub@redhat.com>

PR c++/97145
* constexpr.c (cxx_eval_builtin_function_call): Return void_node for
calls to __sanitize_ptr_{sub,cmp} builtins.

* g++.dg/asan/pr97145.C: New test.

libstdc++: Fix out-of-bounds string_view access in filesystem::path [PR 97167]

libstdc++-v3/ChangeLog:

PR libstdc++/97167
* src/c++17/fs_path.cc (path::_Parser::root_path()): Check
for empty string before inspecting the first character.
* testsuite/27_io/filesystem/path/append/source.cc: Append
empty string_view to path.

analyzer: add -fdump-analyzer-json

I've found this useful for debugging state explosions in the analyzer.

gcc/analyzer/ChangeLog:
* analysis-plan.cc: Include "json.h".
* analyzer.opt (fdump-analyzer-json): New.
* call-string.cc: Include "json.h".
(call_string::to_json): New.
* call-string.h (call_string::to_json): New decl.
* checker-path.cc: Include "json.h".
* constraint-manager.cc: Include "json.h".
(equiv_class::to_json): New.
(constraint::to_json): New.
(constraint_manager::to_json): New.
* constraint-manager.h (equiv_class::to_json): New decl.
(constraint::to_json): New decl.
(constraint_manager::to_json): New decl.
* diagnostic-manager.cc: Include "json.h".
(saved_diagnostic::to_json): New.
(diagnostic_manager::to_json): New.
* diagnostic-manager.h (saved_diagnostic::to_json): New decl.
(diagnostic_manager::to_json): New decl.
* engine.cc: Include "json.h", <zlib.h>.
(exploded_node::status_to_str): New.
(exploded_node::to_json): New.
(exploded_edge::to_json): New.
(exploded_graph::to_json): New.
(dump_analyzer_json): New.
(impl_run_checkers): Call it.
* exploded-graph.h (exploded_node::status_to_str): New decl.
(exploded_node::to_json): New.
(exploded_edge::to_json): New.
(exploded_graph::to_json): New.
* pending-diagnostic.cc: Include "json.h".
* program-point.cc: Include "json.h".
(program_point::to_json): New.
* program-point.h (program_point::to_json): New decl.
* program-state.cc: Include "json.h".
(extrinsic_state::to_json): New.
(sm_state_map::to_json): New.
(program_state::to_json): New.
* program-state.h (extrinsic_state::to_json): New decl.
(sm_state_map::to_json): New decl.
(program_state::to_json): New decl.
* region-model-impl-calls.cc: Include "json.h".
* region-model-manager.cc: Include "json.h".
* region-model-reachability.cc: Include "json.h".
* region-model.cc: Include "json.h".
* region-model.h (svalue::to_json): New decl.
(region::to_json): New decl.
* region.cc: Include "json.h".
(region::to_json: New.
* sm-file.cc: Include "json.h".
* sm-malloc.cc: Include "json.h".
* sm-pattern-test.cc: Include "json.h".
* sm-sensitive.cc: Include "json.h".
* sm-signal.cc: Include "json.h".
(signal_delivery_edge_info_t::to_json): New.
* sm-taint.cc: Include "json.h".
* sm.cc: Include "diagnostic.h", "tree-diagnostic.h", and
"json.h".
(state_machine::state::to_json): New.
(state_machine::to_json): New.
* sm.h (state_machine::state::to_json): New.
(state_machine::to_json): New.
* state-purge.cc: Include "json.h".
* store.cc: Include "json.h".
(binding_key::get_desc): New.
(binding_map::to_json): New.
(binding_cluster::to_json): New.
(store::to_json): New.
* store.h (binding_key::get_desc): New decl.
(binding_map::to_json): New decl.
(binding_cluster::to_json): New decl.
(store::to_json): New decl.
* supergraph.cc: Include "json.h".
(supergraph::to_json): New.
(supernode::to_json): New.
(superedge::to_json): New.
* supergraph.h (supergraph::to_json): New decl.
(supernode::to_json): New decl.
(superedge::to_json): New decl.
* svalue.cc: Include "json.h".
(svalue::to_json): New.

gcc/ChangeLog:
* doc/analyzer.texi (Other Debugging Techniques): Mention
-fdump-analyzer-json.
* doc/invoke.texi (Static Analyzer Options): Add
-fdump-analyzer-json.

bpf: use xBPF signed div, mod insns when available

The 'mod' and 'div' operators in eBPF are unsigned, with no signed
counterpart. xBPF adds two new ALU operations, sdiv and smod, for
signed division and modulus, respectively. Update bpf.md with
'define_insn' blocks for signed div and mod to use them when targetting
xBPF, and add new tests to ensure they are used appropriately.

2020-09-17  David Faust  <david.faust@oracle.com>

gcc/
* config/bpf/bpf.md: Add defines for signed div and mod operators.

gcc/testsuite/
* gcc.target/bpf/diag-sdiv.c: New test.
* gcc.target/bpf/diag-smod.c: New test.
* gcc.target/bpf/xbpf-sdiv-1.c: New test.
* gcc.target/bpf/xbpf-smod-1.c: New test.

compiler: call runtime.eqtype for non-interface type switch on aix

All type switch clauses must call runtime.eqtype if the linker isn't
able to merge type descriptors pointers. Previously, only interface-type
clauses were doing it.

Updates golang/go#39276

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/255202

libgomp.fortran/pr66199-5.f90: Make stop codes unique

libgomp/ChangeLog:

PR fortran/95654
* testsuite/libgomp.fortran/pr66199-5.f90: Make stop codes unique.

libstdc++: Fix overflow handling in std::align

libstdc++-v3/ChangeLog:

* include/bits/align.h (align): Fix overflow handling.
* testsuite/20_util/align/3.cc: New test.

c++: fix injected friend of template class

In working on fixing hiddenness, I discovered some suspicious code in
template instantiation.  I suspect it dates from when we didn't do the
hidden friend injection thing at all.  The xreftag finds the same
class, but makes it visible to name lookup.  Which is wrong.
hurrah, fixing a bug by deleting code!

gcc/cp/
* pt.c (instantiate_class_template_1): Do not repush and unhide
injected friend.
gcc/testsuite/
* g++.old-deja/g++.pt/friend34.C: Check injected friend is still
invisible.

libstdc++: Introduce new headers for C++20 ranges components

This introduces two new headers:

<bits/ranges_base.h> defines the minimal components needed
for using C++20 ranges (customization point objects such as
std::ranges::begin, concepts such as std::ranges::range, etc.)

<bits/ranges_util.h> includes <bits/ranges_base.h> and additionally
defines subrange, which is needed by <bits/ranges_algo.h>.

Most of the content of <bits/ranges_base.h> was previously defined in
<bits/range_access.h>, but a few pieces were only defined in <ranges>.
This meant the entire <ranges> header was needed in <algorithm> and
<memory>, even though they don't use all the range adaptors.

By moving the ranges components out of <bits/range_access.h> that file
is left defining just the contents of [iterator.range] i.e. std::begin,
std::end, std::size etc. and not C++20 ranges components.

For consistency with other C++20 ranges headers, <bits/range_cmp.h> is
renamed to <bits/ranges_cmp.h>.

libstdc++-v3/ChangeLog:

* include/Makefile.am: Add new headers and adjust for renamed
header.
* include/Makefile.in: Regenerate.
* include/bits/iterator_concepts.h: Adjust for renamed header.
* include/bits/range_access.h (ranges::*): Move to new
<bits/ranges_base.h> header.
* include/bits/ranges_algobase.h: Include new <bits/ranges_base.h>
header instead of <ranges>.
* include/bits/ranges_algo.h: Include new <bits/ranges_util.h>
header.
* include/bits/range_cmp.h: Moved to...
* include/bits/ranges_cmp.h: ...here.
* include/bits/ranges_base.h: New header.
* include/bits/ranges_util.h: New header.
* include/experimental/string_view: Include new
<bits/ranges_base.h> header.
* include/std/functional: Adjust for renamed header.
* include/std/ranges (ranges::view_base, ranges::enable_view)
(ranges::dangling, ranges::borrowed_iterator_t): Move to new
<bits/ranges_base.h> header.
(ranges::view_interface, ranges::subrange)
(ranges::borrowed_subrange_t): Move to new <bits/ranges_util.h>
header.
* include/std/span: Include new <bits/ranges_base.h> header.
* include/std/string_view: Likewise.
* testsuite/24_iterators/back_insert_iterator/pr93884.cc: Add
missing <ranges> header.
* testsuite/24_iterators/front_insert_iterator/pr93884.cc:
Likewise.

testsuite: Prune more output in timevar1.C.

gcc/testsuite/ChangeLog:

* g++.dg/ext/timevar1.C: Also prune N%.

testsuite: Prune more output in timevar2.C.

gcc/testsuite/ChangeLog:

* g++.dg/ext/timevar2.C: Also prune N%.

switch lowering: limit number of cluster attemps

gcc/ChangeLog:

PR tree-optimization/96979
* doc/invoke.texi: Document new param max-switch-clustering-attempts.
* params.opt: Add new parameter.
* tree-switch-conversion.c (jump_table_cluster::find_jump_tables):
Limit number of attempts.
(bit_test_cluster::find_bit_tests): Likewise.

gcc/testsuite/ChangeLog:

PR tree-optimization/96979
* g++.dg/tree-ssa/pr96979.C: New test.

IBM Z: Try to make use of load-and-test instructions

This patch enables a peephole2 optimization which transforms a load of
constant zero into a temporary register which is then finally used to
compare against a floating-point register of interest into a single load
and test instruction.  However, the optimization is only applied if both
registers are dead afterwards and if we test for (in)equality only.
This is relaxed in case of fast math.

This is a follow up to PR88856.

gcc/ChangeLog:

* config/s390/s390.md ("*cmp<mode>_ccs_0", "*cmp<mode>_ccz_0",
"*cmp<mode>_ccs_0_fastmath"): Basically change "*cmp<mode>_ccs_0" into
"*cmp<mode>_ccz_0" and for fast math add "*cmp<mode>_ccs_0_fastmath".

gcc/testsuite/ChangeLog:

* gcc.target/s390/load-and-test-fp-1.c: Change test to include all
possible combinations of dead/live registers and comparisons (equality,
relational).
* gcc.target/s390/load-and-test-fp-2.c: Same as load-and-test-fp-1.c
but for fast math.
* gcc.target/s390/load-and-test-fp.h: New test included by
load-and-test-fp-{1,2}.c.

[libgomp, nvptx] Print error log for link error

By running libgomp test-case libgomp.c/target-28.c with GOMP_NVPTX_PTXRW=w
(using a maintenance patch that adds support for this env var), we dump the
ptx in target-28.exe to file.  By editing one ptx file to rename
gomp_nvptx_main to gomp_nvptx_main2 in both declaration and call, and
running with GOMP_NVPTX_PTXRW=r, we trigger a link error:
...
$ GOMP_NVPTX_PTXRW=r ./target-28.exe
libgomp: cuLinkComplete error: unknown error
...
The error is somewhat uninformative.

Fix this by dumping the error log returned by the failing cuda call, such
that we have instead:
...
$ GOMP_NVPTX_PTXRW=r ./target-28.exe
libgomp: Link error log error   : \
  Undefined reference to 'gomp_nvptx_main2' in ''
libgomp: cuLinkComplete error: unknown error
...

Build on x86_64 with nvptx accelerator, tested libgomp.

libgomp/ChangeLog:

* plugin/plugin-nvptx.c (link_ptx): Print elog if cuLinkComplete call
fails.

AArch64: Implement missing vcls intrinsics on unsigned types

This patch implements some missing intrinsics that perform a CLS on unsigned SIMD types.

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

gcc/
PR target/71233
* config/aarch64/arm_neon.h (vcls_u8, vcls_u16, vcls_u32,
vclsq_u8, vclsq_u16, vclsq_u32): Define.

gcc/testsuite/
PR target/71233
* gcc.target/aarch64/simd/vcls_unsigned_1.c: New test.

AArch64: Implement missing vceq*_p* intrinsics

This patch implements some missing vceq* intrinsics on poly types.
The behaviour is to produce the appropriate CMEQ instruction as for the unsigned types.

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

gcc/
PR target/71233
* config/aarch64/arm_neon.h (vceqq_p64, vceqz_p64, vceqzq_p64): Define.

gcc/testsuite/

PR target/71233
* gcc.target/aarch64/simd/vceq_poly_1.c: New test.

AArch64: Implement poly-type vadd intrinsics

This implements the vadd[p]_p* intrinsics.
In terms of functionality they are aliases of veor operations on the relevant unsigned types.

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

gcc/
PR target/71233
* config/aarch64/arm_neon.h (vadd_p8, vadd_p16, vadd_p64, vaddq_p8,
vaddq_p16, vaddq_p64, vaddq_p128): Define.

gcc/testsuite/
PR target/71233
* gcc.target/aarch64/simd/vadd_poly_1.c: New test.

gcov: fix TOPN streaming from shared libraries

Before the change gcc did not stream correctly TOPN counters
if counters belonged to a non-local shared object.

As a result zero-section optimization generated TOPN sections
in a form not recognizable by '__gcov_merge_topn'.

The problem happens because in a case of multiple shared objects
'__gcov_merge_topn' function is present in address space multiple
times (once per each object).

The fix is to never rely on function address and predicate on TOPN
counter types.

libgcc/ChangeLog:

PR gcov-profile/96913
* libgcov-driver.c (write_one_data): Avoid function pointer
comparison in TOP streaming decision.

Fix some typos

2020-09-22  Jakub Jelinek  <jakub@redhat.com>

gcc/
* params.opt (--param=modref-max-tests=): Fix typo in help text:
perofmed -> performed.
* common.opt: Fix typo: incrmeental -> incremental.
* ipa-modref.c: Fix typos: recroding -> recording, becaue -> because,
analsis -> analysis.
(class modref_summaries): Fix typo: betweehn -> between.
(analyze_call): Fix typo: calle -> callee.
(read_modref_records): Fix typo: expcted -> expected.
(pass_ipa_modref::execute): Fix typo: calle -> callee.
gcc/c-family/
* c.opt (Wbuiltin-declaration-mismatch): Fix typo in variable name:
warn_builtin_declaraion_mismatch -> warn_builtin_declaration_mismatch.

ipa: Fix up ipa modref option help texts

This fixes
FAIL: compiler driver --help=common option(s): "^ +-.*[^:.]$" absent from output: "  --param=modref-max-tests=   Maximum number of tests perofmed by modref query"
FAIL: compiler driver --help=optimizers option(s): "^ +-.*[^:.]$" absent from output: "  -fipa-modref                Perform interprocedural modref analysis"

2020-09-22  Jakub Jelinek  <jakub@redhat.com>

* common.opt (-fipa-modref): Add dot at the end of option help.
* params.opt (--param=modref-max-tests=): Likewise.

libstdc++: Use correct argument type for __use_alloc, again [PR 96803]

While backporting 5494edae83ad33c769bd1ebc98f0c492453a6417 I noticed
that it's still not correct. I made the allocator-extended constructor
use the right type for the uses-allocator construction detection, but I
used an rvalue when it should be a const lvalue.

This should fix it properly this time.

libstdc++-v3/ChangeLog:

PR libstdc++/96803
* include/std/tuple
(_Tuple_impl(allocator_arg_t, Alloc, const _Tuple_impl<U...>&)):
Use correct value category in __use_alloc call.
* testsuite/20_util/tuple/cons/96803.cc: Check with constructors
that require correct value category to be used.

Add no strict aliasing to m128-check.h

* gcc.target/i386/m128-check.h: Add no-strict-aliasing

libstdc++: Remove overzealous static_asserts from std::span

For a span with statically empty extent, we currently model the
preconditions of front(), back(), and operator[] as if they are
mandates, by using a static_assert to verify that extent != 0.  This
causes us to reject valid programs that would instantiate these member
functions and at runtime never call them.

Since they are already followed by more general runtime asserts, this
patch just removes these static_asserts altogether,

libstdc++-v3/ChangeLog:

* include/std/span (span::front): Remove static_assert.
(span::back): Likewise.
(span::operator[]): Likewise.
* testsuite/23_containers/span/back_neg.cc: Rewrite to verify
that we check the preconditions of back() only when it's called.
* testsuite/23_containers/span/front_neg.cc: Likewise for
front().
* testsuite/23_containers/span/index_op_neg.cc: Likewise for
operator[].

libstdc++: Fix division by zero in std::sample

This fixes a division by zero in the selection-sampling std::__sample
overload when the input range is empty (and hence __unsampled_sz is 0).

libstdc++-v3/ChangeLog:

* include/bits/stl_algo.h (__sample): Exit early when the
input range is empty.
* testsuite/25_algorithms/sample/3.cc: New test.

libstdc++: Mark some more algorithms constexpr for C++20

As per P0202.

libstdc++-v3/ChangeLog:

* include/bits/stl_algo.h (for_each_n): Mark constexpr for C++20.
(search): Likewise for the overload that takes a searcher.
* testsuite/25_algorithms/for_each/constexpr.cc: Test constexpr
std::for_each_n.
* testsuite/25_algorithms/search/constexpr.cc: Test constexpr
std::search overload that takes a searcher.

Daily bump.

compiler: finalize methods for type aliases of struct types

Previously we would finalize the methods of the alias type itself, but
since its a type alias we really need to finalize the methods of the
aliased type.

Also, handle method expressions of unnamed struct types.

Test case is https://golang.org/cl/251168.

Fixes golang/go#38125

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/251279

analyzer: fix ICE on bogus decl of memset [PR97130]

Verify that arguments are pointers before calling handling code
that calls deref_rvalue on them.

gcc/analyzer/ChangeLog:
PR analyzer/97130
* region-model-impl-calls.cc (call_details::get_arg_type): New.
* region-model.cc (region_model::on_call_pre): Check that the
initial arg is a pointer before calling impl_call_memset and
impl_call_strlen.
* region-model.h (call_details::get_arg_type): New decl.

gcc/testsuite/ChangeLog:
PR analyzer/97130
* gcc.dg/analyzer/pr97130.c: New test.

analyzer: decls are not on the heap

Whilst debugging the remaining state explosion in PR analyzer/93355
I noticed that half of the states at an exploding program point had:
  'malloc': {'&buf': 'non-heap'}
whereas the other half didn't, presumably depending on whether the path
to each enode had used this local buffer:
  char buf[400];

This patch tweaks malloc_state_machine::get_default_state to be smarter
about this, so that we can implicitly treat pointers to decls as
non-heap, preventing pointless differences between sm_state_map
instances.  With that, all of the states in question have equal (empty)
malloc sm-state - though the state explosion continues for other reasons.

gcc/analyzer/ChangeLog:
PR analyzer/93355
* sm-malloc.cc (malloc_state_machine::get_default_state): Look at
the base region when considering pointers.  Treat pointers to
decls as being non-heap.

libstdc++: Use __builtin_expect in __glibcxx_assert

libstdc++-v3/ChangeLog:

* include/bits/c++config (__replacement_assert): Add noreturn
attribute.
(__glibcxx_assert_impl): Use __builtin_expect to hint that the
assertion is expected to pass.

libstdc++: Fix constraints for drop_view::begin() const [LWG 3482]

libstdc++-v3/ChangeLog:

* include/std/ranges (drop_view::begin()): Adjust constraints
to match the correct condition for O(1) ranges::next (LWG 3482).
* testsuite/std/ranges/adaptors/drop.cc: Check that iterator is
cached for non-sized_range.

go/internal/gccgoimporter: recognize aixbigafMagic archives

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/255201

c++: DR 1722: Make lambda to function pointer conv noexcept [PR90583]

DR 1722 clarifies that the conversion function from lambda to pointer to
function should be noexcept(true).

gcc/cp/ChangeLog:

PR c++/90583
DR 1722
* lambda.c (maybe_add_lambda_conv_op): Mark the conversion function
as noexcept.

gcc/testsuite/ChangeLog:

PR c++/90583
DR 1722
* g++.dg/cpp0x/lambda/lambda-conv14.C: New test.

c++: Implement -Wctad-maybe-unsupported.

I noticed that clang++ has this CTAD warning and thought that it might
be useful to have it.  From clang++: "Some style guides want to allow
using CTAD only on types that "opt-in"; i.e. on types that are designed
to support it and not just types that *happen* to work with it."

So this warning warns when CTAD deduced a type, but the type does not
define any deduction guides.  In that case CTAD worked only because the
compiler synthesized the implicit deduction guides.  That might not be
intended.

It can be suppressed by adding a deduction guide that will never be
considered:

  struct allow_ctad_t;
  template <typename T> struct S { S(T) {} };
  S(allow_ctad_t) -> S<void>;

This warning is off by default.  It doesn't warn when the type comes
from a system header unless -Wsystem-headers.

gcc/c-family/ChangeLog:

* c.opt (Wctad-maybe-unsupported): New option.

gcc/cp/ChangeLog:

* pt.c (deduction_guides_for): Add a bool parameter.  Set it.
(do_class_deduction): Warn when CTAD succeeds but the type doesn't
have any explicit deduction guides.

gcc/ChangeLog:

* doc/invoke.texi: Document -Wctad-maybe-unsupported.

gcc/testsuite/ChangeLog:

* g++.dg/warn/Wctad-maybe-unsupported.C: New test.
* g++.dg/warn/Wctad-maybe-unsupported2.C: New test.
* g++.dg/warn/Wctad-maybe-unsupported3.C: New test.
* g++.dg/warn/Wctad-maybe-unsupported.h: New file.

libgo: don't put golang.org packages in zstdpkglist.go

This ensures that internal/goroot.IsStandardPackage does not treat
golang.org packages as being in the standard library.

For golang/go#41368
Fixes golang/go#41499

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/256319

Avoid incidental failures due to implicit attribute access.

gcc/testsuite/ChangeLog:

PR c/50584
* gcc.dg/ipa/ipa-sra-1.c: Use a plain pointer for argv instead of array.
* gcc.dg/ipa/ipa-sra-12.c: Same.
* gcc.dg/ipa/ipa-sra-13.c: Same.
* gcc.dg/ipa/ipa-sra-14.c: Same.
* gcc.dg/ipa/ipa-sra-15.c: Same.

PR fortran/90903 [part2] - Add runtime checking for the MVBITS intrinsic

Implement inline expansion of the intrinsic elemental subroutine MVBITS
with optional runtime checks for valid argument range.

gcc/fortran/ChangeLog:

* iresolve.c (gfc_resolve_mvbits): Remove unneeded conversion of
FROMPOS, LEN and TOPOS arguments to fit a C int.
* trans-intrinsic.c (gfc_conv_intrinsic_mvbits): Add inline
expansion of MVBITS intrinsic elemental subroutine and add code
for runtime argument checking.
(gfc_conv_intrinsic_subroutine): Recognise MVBITS intrinsic, but
defer handling to gfc_trans_call.
* trans-stmt.c (replace_ss):
(gfc_trans_call): Adjust to handle inline expansion, scalarization
of intrinsic subroutine MVBITS in gfc_conv_intrinsic_mvbits.
* trans.h (gfc_conv_intrinsic_mvbits): Add prototype for
gfc_conv_intrinsic_mvbits.

gcc/testsuite/ChangeLog:

* gfortran.dg/check_bits_2.f90: New test.

Co-authored-by: Paul Thomas <pault@gcc.gnu.org>

libsanitizer: rename new libbacktrace symbols

* libbacktrace/backtrace-rename.h (backtrace_uncompress_lzma):
Define.
(backtrace_syminfo_to_full_callback): Define.
(backtrace_syminfo_to_full_error_callback): Define.

c++: ts_lambda is not needed

We don't need ts_lambda, as IDENTIFIER_LAMBDA_P is sufficient.  Killed thusly.

gcc/cp/
* decl.c (xref_tag_1): Use IDENTIFIER_LAMBDA_P to detect lambdas.
* lambda.c (begin_lambda_type): Use ts_current to push the tag.
* name-lookup.h (enum tag_scope): Drop ts_lambda.

Darwin, testsuite : Skip a test that requires ELF.

The symver support is only available to ELF targets.

gcc/testsuite/ChangeLog:

* gcc.dg/ipa/symver1.c: Skip for Darwin.

[arm] gcc.target/arm/cs*: Use dg-add-options arm_arch_v8_1m_main

These testcases need thumb mode, which may not be the default.

Using dg-add-options arm_arch_v8_1m_main ensures that -mthumb is used
and makes the test pass in more configurations.

2020-09-21  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/testsuite/
* gcc.target/arm/csinc-1.c: Use dg-add-options
arm_arch_v8_1m_main.
* gcc.target/arm/csinv-1.c: Likewise.
* gcc.target/arm/csneg.c: Likewise.

tree-optimization/97139 - fix BB SLP live lane extraction

This fixes SLP live lane extraction with pattern stmts.

2020-09-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/97139
* tree-vect-slp.c (vect_bb_slp_mark_live_stmts): Only mark the
pattern root, track visited vectorized stmts.

* gcc.dg/vect/pr97139.c: New testcase.

c++: Detect deduction guide redeclaration [PR97099]

[temp.deduct.guide]p3: Two deduction guide declarations in the same
translation unit for the same class template shall not have equivalent
parameter-declaration-clauses.

So let's detect that.

gcc/cp/ChangeLog:

PR c++/97099
* decl.c (redeclaration_error_message): Detect a redeclaration of
deduction guides.

gcc/testsuite/ChangeLog:

PR c++/97099
* g++.dg/cpp1z/class-deduction74.C: New test.

libstdc++: Relax constraints on transform_view and elements_view iterators

libstdc++-v3/ChangeLog:

* include/std/ranges (transform_view, elements_view): Relax
constraints on operator- for iterators, as per LWG 3483.
* testsuite/std/ranges/adaptors/elements.cc: Check that we
can take the difference of two iterators from a non-random
access range.
* testsuite/std/ranges/adaptors/transform.cc: Likewise.

libstdc++: Make std::assume_aligned a constexpr function [PR 97132]

The cast from void* to T* in std::assume_aligned is not valid in a
constexpr function. The optimization hint is redundant during constant
evaluation anyway (the compiler can see the object and knows its
alignment). Simply return the original pointer without applying the
__builtin_assume_aligned hint to it when doing constant evaluation.

This change also removes the preprocessor branch that works around
uintptr_t not being available. We already assume that type is present
elsewhere in the library.

libstdc++-v3/ChangeLog:

PR libstdc++/97132
* include/bits/align.h (align) [!_GLIBCXX_USE_C99_STDINT_TR1]:
Remove unused code.
(assume_aligned): Do not use __builtin_assume_aligned during
constant evaluation.
* testsuite/20_util/assume_aligned/1.cc: Improve test.
* testsuite/20_util/assume_aligned/97132.cc: New test.

Fix fallout from Support new mallinfo2 function.

2020-09-21  Jakub Jelinek  <jakub@redhat.com>

* configure.ac: Use mallinfo mallinfo2 as first operand of
gcc_AC_CHECK_DECLS rather than [mallinfo, mallinfo2].
* configure: Regenerated.
* config.in: Regenerated.

aarch64: Do not alter value on a force_reg returned rtx expanding __jcvt

2020-09-17  Andrea Corallo  <andrea.corallo@arm.com>

* config/aarch64/aarch64-builtins.c
(aarch64_general_expand_builtin): Use expand machinery not to
alter the value of an rtx returned by force_reg.

tree-optimization/97135 - fix dependence check in store-motion

The following fixes a dependence check where in the particular place
we cannot ignore self-dependences.

2020-09-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/97135
* tree-ssa-loop-im.c (sm_seq_push_down): Do not ignore
self-dependences.

* gcc.dg/torture/pr97135.c: New testcase.

Fix ICE in tree-switch-conversion.

With SVE we can end up with:
switch (POLY_INT_CST [2, 2]) <default: <L2> [INV], case 2: <L3> [INV], case 4: <L3> [INV]>
which is fine to expand and we can remove the assert.

gcc/ChangeLog:

PR tree-optimization/96915
* tree-switch-conversion.c (switch_conversion::expand): Accept
also integer constants.

gcc/testsuite/ChangeLog:

PR tree-optimization/96915
* gcc.target/aarch64/sve/pr96915.c: New test.

POLY_INT_CST: remove extra space in dump

gcc/ChangeLog:

* print-tree.c (print_node): Remove extra space.

mklog.py: Parse first 10 lines for PR/DR number

contrib/ChangeLog:

* mklog.py: Parse first 10 lines for PR/DR number
not only the first line.

aarch64: Fix ICE on fpsr fpcr getters [PR96968]

gcc/ChangeLog

2020-09-14  Andrea Corallo  <andrea.corallo@arm.com>

PR target/96968
* config/aarch64/aarch64-builtins.c
(aarch64_expand_fpsr_fpcr_setter): Fix comment nit.
(aarch64_expand_fpsr_fpcr_getter): New function, expand these
getters using expand_insn machinery.
(aarch64_general_expand_builtin): Make use of.

gcc/testsuite/ChangeLog

2020-09-14  Andrea Corallo  <andrea.corallo@arm.com>

PR target/96968
* gcc.target/aarch64/pr96968.c: New test.

Use ONE_? macros.

gcc/ChangeLog:

* ggc-common.c (ggc_rlimit_bound): Use ONE_? macro.
(ggc_min_expand_heuristic): Likewise.
(ggc_min_heapsize_heuristic): Likewise.
* ggc-page.c (ggc_collect): Likewise.
* system.h (ONE_G): Likewise.

Use SIZE_AMOUNT macro for GGC memory allocation numbers.

gcc/ChangeLog:

* ggc-common.c (ggc_prune_overhead_list): Use SIZE_AMOUNT.
* ggc-page.c (release_pages): Likewise.
(ggc_collect): Likewise.
(ggc_trim): Likewise.
(ggc_grow): Likewise.
* timevar.c (timer::print): Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/ext/timevar1.C: Prune more possible number values.
* g++.dg/ext/timevar2.C: Likewise.

Support new mallinfo2 function.

gcc/ChangeLog:

* config.in: Regenerate.
* configure: Likewise.
* configure.ac: Detect for mallinfo2.
* ggc-common.c (defined): Use it.
* system.h: Handle also HAVE_MALLINFO2.

Daily bump.

libstdc++: Fix noexcept-specifier for std::bind_front [PR 97101]

libstdc++-v3/ChangeLog:

PR libstdc++/97101
* include/std/functional (bind_front): Fix order of parameters
in is_nothrow_constructible_v specialization.
* testsuite/20_util/function_objects/bind_front/97101.cc: New test.

Fix linkage with -nodefaultlibs option.

2020-09-20  John David Anglin  < danglin@gcc.gnu.org>

gcc/ChangeLog
* config/pa/pa-hpux11.h (LINK_GCC_C_SEQUENCE_SPEC): Delete.
* config/pa/pa64-hpux.h (LINK_GCC_C_SEQUENCE_SPEC): Likewise.
(ENDFILE_SPEC): Link with libgcc_stub.a and mill.a.
* config/pa/pa32-linux.h (ENDFILE_SPEC): Link with libgcc.a.

c++: Add test for PR90199.

Fixed by r11-2998, which fixed this ICE too.

gcc/testsuite/ChangeLog:

PR c++/90199
* g++.dg/cpp1y/constexpr-90199.C: New test.

Fix ICE during dumping in ipa-modref.

2020-09-20  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref.c (dump_lto_records): Fix ICE.

New modref/ipa_modref optimization passes

2020-09-19  David Cepelik  <d@dcepelik.cz>
    Jan Hubicka  <hubicka@ucw.cz>

* Makefile.in: Add ipa-modref.c and ipa-modref-tree.c.
* alias.c: (reference_alias_ptr_type_1): Export.
* alias.h (reference_alias_ptr_type_1): Declare.
* common.opt (fipa-modref): New.
* gengtype.c (open_base_files): Add ipa-modref-tree.h and ipa-modref.h
* ipa-modref-tree.c: New file.
* ipa-modref-tree.h: New file.
* ipa-modref.c: New file.
* ipa-modref.h: New file.
* lto-section-in.c (lto_section_name): Add ipa_modref.
* lto-streamer.h (enum lto_section_type): Add LTO_section_ipa_modref.
* opts.c (default_options_table): Enable ipa-modref at -O1+.
* params.opt (-param=modref-max-bases, -param=modref-max-refs,
-param=modref-max-tests): New params.
* passes.def: Schedule pass_modref and pass_ipa_modref.
* timevar.def (TV_IPA_MODREF): New timevar.
(TV_TREE_MODREF): New timevar.
* tree-pass.h (make_pass_modref): Declare.
(make_pass_ipa_modref): Declare.
* tree-ssa-alias.c (dump_alias_stats): Include ipa-modref-tree.h
and ipa-modref.h
(alias_stats): Add modref_use_may_alias, modref_use_no_alias,
modref_clobber_may_alias, modref_clobber_no_alias, modref_tests.
(dump_alias_stats): Dump new stats.
(nonoverlapping_array_refs_p): Fix formating.
(modref_may_conflict): New function.
(ref_maybe_used_by_call_p_1): Use it.
(call_may_clobber_ref_p_1): Use it.
(call_may_clobber_ref_p): Update.
(stmt_may_clobber_ref_p_1): Update.
* tree-ssa-alias.h (call_may_clobber_ref_p_1): Update.

Daily bump.

Extend -Warray-bounds to detect out-of-bounds accesses to array parameters.

gcc/ChangeLog:

PR middle-end/82608
PR middle-end/94195
PR c/50584
PR middle-end/84051
* gimple-array-bounds.cc (get_base_decl): New function.
(get_ref_size): New function.
(trailing_array): New function.
(array_bounds_checker::check_array_ref): Call them.  Handle arrays
declared in function parameters.
(array_bounds_checker::check_mem_ref):  Same.  Handle references to
dynamically allocated arrays.

gcc/testsuite/ChangeLog:

PR middle-end/82608
PR middle-end/94195
PR c/50584
PR middle-end/84051
* c-c++-common/Warray-bounds.c: Adjust.
* gcc.dg/Wbuiltin-declaration-mismatch-9.c: Adjust.
* gcc.dg/Warray-bounds-63.c: New test.
* gcc.dg/Warray-bounds-64.c: New test.
* gcc.dg/Warray-bounds-65.c: New test.
* gcc.dg/Warray-bounds-66.c: New test.
* gcc.dg/Warray-bounds-67.c: New test.

Extend -Wstringop-overflow to detect out-of-bounds accesses to array parameters.

gcc/ChangeLog:

PR c/50584
* builtins.c (warn_for_access): Add argument.  Distinguish between
reads and writes.
(check_access): Add argument.  Distinguish between reads and writes.
(gimple_call_alloc_size): Set range even on failure.
(gimple_parm_array_size): New function.
(compute_objsize): Call it.
(check_memop_access): Pass check_access an additional argument.
(expand_builtin_memchr, expand_builtin_strcat): Same.
(expand_builtin_strcpy, expand_builtin_stpcpy_1): Same.
(expand_builtin_stpncpy, check_strncat_sizes): Same.
(expand_builtin_strncat, expand_builtin_strncpy): Same.
(expand_builtin_memcmp): Same.
* builtins.h (compute_objsize): Declare a new overload.
(gimple_parm_array_size): Declare.
(check_access): Add argument.
* calls.c (append_attrname): Simplify.
(maybe_warn_rdwr_sizes): Handle internal attribute access.
* tree-ssa-uninit.c (maybe_warn_pass_by_reference): Avoid adding
quotes.

gcc/testsuite/ChangeLog:

PR c/50584
* c-c++-common/Wsizeof-pointer-memaccess1.c: Disable new expected
warnings.
* g++.dg/ext/attr-access.C: Update text of expected warnings.
* gcc.dg/Wstringop-overflow-23.c: Same.
* gcc.dg/Wstringop-overflow-24.c: Same.
* gcc.dg/attr-access-none.c: Same.
* gcc.dg/dfp/composite-type.c: Prune expected warnings.
* gcc.dg/torture/pr57147-1.c: Add a member to an otherwise empty
struct to avoid a warning.
* gcc.dg/torture/pr57147-3.c: Same.
* gcc.dg/Warray-bounds-30.c: Adjust.
* gcc.dg/attr-access-none.c: Same.
* gcc.dg/Wstringop-overflow-40.c: New test.
* gcc.dg/attr-access-2.c: New test.

Make use of new attribute access infrastructure in -Wuninitialized (PR 50584).

gcc/ChangeLog:

* tree-ssa-uninit.c (maybe_warn_pass_by_reference): Handle attribute
access internal representation of arrays.

gcc/testsuite/ChangeLog:

* gcc.dg/uninit-37.c: New test.

Infrastructure & C front end changes for array parameter checking (PR c/50584).

gcc/ChangeLog:

PR c/50584
* attribs.c (decl_attributes): Also pass decl along with type
attributes to handlers.
(init_attr_rdwr_indices): Change second argument to attribute chain.
Handle internal attribute representation in addition to external.
(get_parm_access): New function.
(attr_access::to_internal_string): Define new member function.
(attr_access::to_external_string): Define new member function.
(attr_access::vla_bounds): Define new member function.
* attribs.h (struct attr_access): Declare new members.
(attr_access::from_mode_char): Define new member function.
(get_parm_access): Declare new function.
* calls.c (initialize_argument_information): Pass function type
attributes to init_attr_rdwr_indices.
* doc/invoke.texi (-Warray-parameter, -Wvla-parameter): Document.
* tree-pretty-print.c (dump_generic_node): Correct handling of
qualifiers.
* tree-ssa-uninit.c (maybe_warn_pass_by_reference): Same.
* tree.h (access_mode): Add new enumerator.

gcc/c-family/ChangeLog:

PR c/50584
* c-attribs.c (c_common_attribute_table): Add "arg spec" attribute.
(handle_argspec_attribute): New function.
(get_argument, get_argument_type): New functions.
(append_access_attrs): Add overload.  Handle internal attribute
representation in addition to external.
(handle_access_attribute): Handle internal attribute representation
in addition to external.
(build_attr_access_from_parms): New function.

gcc/c-family/ChangeLog:

PR c/50584
* c-common.h (warn_parm_array_mismatch): Declare new function.
(has_attribute): Move declaration of an existing function.
(build_attr_access_from_parms): Declare new function.
* c-warn.c (parm_array_as_string): Define new function.
(plus_one):  Define new function.
(warn_parm_ptrarray_mismatch): Define new function.
(warn_parm_array_mismatch):  Define new function.
(vla_bound_parm_decl): New function.
* c.opt (-Warray-parameter, -Wvla-parameter): New options.
* c-pretty-print.c (pp_c_type_qualifier_list): Don't print array type
qualifiers here...
(c_pretty_printer::direct_abstract_declarator): ...but instead print
them in brackets here.  Also print [static].  Strip extraneous
expressions from VLA bounds.

gcc/c/ChangeLog:

PR c/50584
* c-decl.c (lookup_last_decl): Define new function.
(c_decl_attributes): Call it.
(start_decl): Add argument and use it.
(finish_decl): Call build_attr_access_from_parms and decl_attributes.
(get_parm_array_spec): Define new function.
(push_parm_decl): Call get_parm_array_spec.
(start_function): Call warn_parm_array_mismatch.  Build attribute
access and add it to current function.
* c-parser.c (c_parser_declaration_or_fndef): Diagnose mismatches
in forms of array parameters.
* c-tree.h (start_decl): Add argument.

gcc/testsuite/ChangeLog:

PR c/50584
* gcc.dg/attr-access-read-write-2.c: Adjust text of expected message.
* c-c++-common/Warray-bounds-6.c: Correct C++ declaration, adjust
text of expected diagnostics.
* gcc.dg/Wbuiltin-declaration-mismatch-9.c: Prune expected warning.
* gcc.dg/Warray-parameter-2.c: New test.
* gcc.dg/Warray-parameter-3.c: New test.
* gcc.dg/Warray-parameter-4.c: New test.
* gcc.dg/Warray-parameter-5.c: New test.
* gcc.dg/Warray-parameter.c: New test.
* gcc.dg/Wvla-parameter-2.c: New test.
* gcc.dg/Wvla-parameter-3.c: New test.
* gcc.dg/Wvla-parameter.c: New test.

Change C front end to emit structured loop and switch tree nodes.

2020-08-12  Sandra Loosemore  <sandra@codesourcery.com>

gcc/c
* c-decl.c (c_break_label, c_cont_label): Delete, and replace
with...
(in_statement): New.
(start_function): Adjust for above change.
(c_push_function_context, c_pop_function_context): Likewise.
* c-lang.h (struct language_function): Likewise.
* c-objc-common.h (LANG_HOOKS_BLOCK_MAY_FALLTHRU): Define.
* c-parser.c (objc_foreach_break_label, objc_foreach_continue_label):
New.
(c_parser_statement_after_labels): Adjust calls to c_finish_bc_stmt.
(c_parser_switch_statement): Adjust break/switch context handling
and calls to renamed functions.
(c_parser_while_statement): Adjust break/switch context handling and
build a WHILE_STMT.
(c_parser_do_statement): Ditto, with DO_STMT respectively.
(c_parser_for_statement): Ditto, with FOR_STMT respectively.
(c_parser_omp_for_loop): Adjust break/switch context handling.
* c-tree.h (c_break_label, c_cont_label): Delete.
(IN_SWITCH_STMT, IN_ITERATION_STMT): Define.
(IN_OMP_BLOCK, IN_OMP_FOR, IN_OBJC_FOREACH): Define.
(in_statement, switch_statement_break_seen_p): Declare.
(c_start_case, c_finish_case): Renamed to...
(c_start_switch, c_finish_switch).
(c_finish_bc_stmt): Adjust arguments.
* c-typeck.c (build_function_call_vec): Don't try to print
statements with %qE format.
(struct c_switch):  Rename switch_expr field to switch_stmt.
Add break_stmt_seen_p field.
(c_start_case): Rename to c_start_switch.  Build a SWITCH_STMT
instead of a SWITCH_EXPR.  Update for changes to struct c_switch.
(do_case): Update for changes to struct c_switch.
(c_finish_case): Rename to c_finish_switch.  Update for changes to
struct c_switch and change of representation from SWITCH_EXPR to
SWITCH_STMT.
(c_finish_loop): Delete.
(c_finish_bc_stmt): Update to reflect changes to break/continue
state representation.  Build a BREAK_STMT or CONTINUE_STMT instead
of a GOTO_EXPR except for objc foreach loops.

gcc/objc
* objc-act.c (objc_start_method_definition): Update to reflect
changes to break/continue state bookkeeping in C front end.

gcc/testsuite/
* gcc.dg/gomp/block-7.c: Update expected error message wording.

Work around bootstrap failure in Fortran front end.

Switching the C++ front end to lower loops the same was as the C front
end triggered this error when bootstrapping the Fortran front end:

/path/to/gcc/fortran/interface.c:3546:12: error: '*new_arg' may be used uninitialized [-Werror=maybe-uninitialized]
 3546 |   new_arg[i]->next = NULL;
      |   ~~~~~~~~~^

Work around this by adding an assertion, which seems appropriate for
documentation and good coding practices anyway.

2020-08-12  Sandra Loosemore  <sandra@codesourcery.com>

gcc/fortran/
* interface.c (gfc_compare_actual_formal): Add assertion after
main processing loop to silence maybe-uninitialized error.

Use C-style loop lowering instead of C++-style.

The C and C++ front ends used to use the same strategy of lowering
loops to gotos with the end test canonicalized to the bottom of the
loop.  In 2014 the C++ front end was changed to emit LOOP_EXPRs
instead (commit 1a45860e7757ee054f6bf98bee4ebe5c661dfb90).

As part of the unification of the C and C++ loop handling, it's
desirable to use the same lowering strategy for both languages.
Applying the C++ strategy to C caused a number of regressions in C
optimization tests, related to flipping the sense of the COND_EXPR for
the exit test and changes in block ordering in the output code.  Many
of these regressions just require updating regexps in the test cases
but a few appear to be genuine optimization failures.  Since it
appears the optimizers handle the C code better than C++ code, let's
go back to using the C strategy for both languages.  The rationale for
the 2014 C++ patch (support for constexpr evaluation) has been solved
in other ways meanwhile.

2020-08-12  Sandra Loosemore  <sandra@codesourcery.com>

gcc/c-family/
* c-gimplify.c (genericize_c_loop): Rewrite to match
c_finish_loop in c-typeck.c.

Move loop and switch tree data structures from cp/ to c-family/.

This patch moves the definitions for DO_STMT, FOR_STMT, WHILE_STMT,
SWITCH_STMT, BREAK_STMT, and CONTINUE_STMT from the C++ front end to
c-family.  This includes the genericizers, pretty-printers, and dump
support as well as the tree definitions and accessors.  Some related
code for OMP_FOR and similar OMP constructs is also moved.

2020-08-12  Sandra Loosemore  <sandra@codesourcery.com>

gcc/c-family/
* c-common.c (c_block_may_fallthrough): New, split from
cxx_block_may_fallthrough in the cp front end.
(c_common_init_ts): Move handling of loop and switch-related
statements here from the cp front end.
* c-common.def (FOR_STMT, WHILE_STMT, DO_STMT): Move here from
cp front end.
(BREAK_STMT, CONTINUE_STMT, SWITCH_STMT): Likewise.
* c-common.h (c_block_may_fallthru): Declare.
(bc_state_t): Move here from cp front end.
(save_bc_state, restore_bc_state): Declare.
(c_genericize_control_stmt): Declare.
(WHILE_COND, WHILE_BODY): Likewise.
(DO_COND, DO_BODY): Likewise.
(FOR_INIT_STMT, FOR_COND, FOR_EXPR, FOR_BODY, FOR_SCOPE): Likewise.
(SWITCH_STMT_COND, SWITCH_STMT_BODY): Likewise.
(SWITCH_STMT_TYPE, SWITCH_STMT_SCOPE): Likewise.
(SWITCH_STMT_ALL_CASES_P, SWITCH_STMT_NO_BREAK_P): Likewise.
(LABEL_DECL_BREAK, LABEL_DECL_CONTINUE): Likewise.
* c-dump.c (dump_stmt): Copy from cp front end.
(c_dump_tree): Move code to handle structured loop and switch
tree nodes here from cp front end.
* c-gimplify.c: Adjust includes.
(enum bc_t, bc_label, begin_bc_block, finish_bc_block): Move from
cp front end.
(save_bc_state, restore_bc_state): New functions using old code
from cp front end.
(get_bc_label, expr_loc_or_loc): Move from cp front end.
(genericize_c_loop): Move from cp front end.
(genericize_for_stmt, genericize_while_stmt): Likewise.
(genericize_do_stmt, genericize_switch_stmt): Likewise.
(genericize_continue_stmt, genericize_break_stmt): Likewise.
(genericize_omp_for_stmt): Likewise.
(c_genericize_control_stmt): New function using code split from
cp front end.
(c_genericize_control_r): New.
(c_genericize): Call walk_tree with c_genericize_control_r.
* c-pretty-print.c (c_pretty_printer::statement): Move code to handle
structured loop and switch tree nodes here from cp front end.

gcc/cp/
* cp-gimplify.c (enum bc_t, bc_label): Move to c-family.
(begin_bc_block, finish_bc_block, get_bc_label): Likewise.
(genericize_cp_loop): Likewise.
(genericize_for_stmt, genericize_while_stmt): Likewise.
(genericize_do_stmt, genericize_switch_stmt): Likewise.
(genericize_continue_stmt, genericize_break_stmt): Likewise.
(genericize_omp_for_stmt): Likewise.
(cp_genericize_r): Call c_genericize_control_stmt instead of
above functions directly.
(cp_genericize): Call save_bc_state and restore_bc_state instead
of manipulating bc_label directly.
* cp-objcp-common.c (cxx_block_may_fallthru): Defer to
c_block_may_fallthru instead of handling SWITCH_STMT here.
(cp_common_init_ts): Move handling of loop and switch-related
statements to c-family.
* cp-tree.def (FOR_STMT, WHILE_STMT, DO_STMT): Move to c-family.
(BREAK_STMT, CONTINUE_STMT, SWITCH_STMT): Likewise.
* cp-tree.h (LABEL_DECL_BREAK, LABEL_DECL_CONTINUE): Likewise.
(WHILE_COND, WHILE_BODY): Likewise.
(DO_COND, DO_BODY): Likewise.
(FOR_INIT_STMT, FOR_COND, FOR_EXPR, FOR_BODY, FOR_SCOPE): Likewise.
(SWITCH_STMT_COND, SWITCH_STMT_BODY): Likewise.
(SWITCH_STMT_TYPE, SWITCH_STMT_SCOPE): Likewise.
(SWITCH_STMT_ALL_CASES_P, SWITCH_STMT_NO_BREAK_P): Likewise.
* cxx-pretty-print.c (cxx_pretty_printer::statement): Move code
to handle structured loop and switch tree nodes to c-family.
* dump.c (cp_dump_tree): Likewise.

gcc/
* doc/generic.texi (Basic Statements): Document SWITCH_EXPR here,
not SWITCH_STMT.
(Statements for C and C++): Rename node to reflect what
the introduction already says about sharing between C and C++
front ends.  Copy-edit and correct documentation for structured
loops and switch.

PR fortran/97036 - [F2018] Allow ELEMENTAL RECURSIVE procedure prefix

gcc/fortran/ChangeLog:

* symbol.c (gfc_check_conflict): Allow ELEMENTAL RECURSIVE
procedure prefix for -std=f2018.

gcc/testsuite/ChangeLog:

* gfortran.dg/pr97036.f90: New test.

x86: Add a testcase for PR target/96861

Add a testcase to verify that -march=skylake-avx512 -mtune=skylake-avx512
generates desired code sequence.

PR target/96861
* gcc.target/i386/pr96861.c: New test.

c++: Fix self-mapping in map_arguments [PR96531, PR97103]

With r10-8077 we stopped passing the argified current_template_parms to
normalize_constraint_expression from finish_nested_requirement, and
instead made map_arguments perform a self-mapping of parameters when
args is NULL.  But we're currently not handling parameter packs and
BOUND_TEMPLATE_TEMPLATE_PARMs properly during this self-mapping, which
leads to ICEs later during satisfaction.

To properly handle self-mapping of a parameter pack, this patch
extends template_parm_to_arg to handle TEMPLATE_PARM_P nodes, and
makes map_arguments use it.  (This change revealed that the call to
template_parm_to_arg in convert_generic_types_to_packs was a no-op
because the argument 't' is never a TREE_LIST, so this patch
additionally removes this call.)

As for bound ttps, map_arguments before r10-8077 would map a
BOUND_TEMPLATE_TEMPLATE_PARM not to itself but to its underlying
TEMPLATE_TEMPLATE_PARM.  We could restore this behavior in
map_arguments, but since a bound ttp is not really a template parameter
it seems better to make keep_template_parm not give us a bound ttp in
the first place.  So this patch makes keep_template_parm return the
underlying ttp when it sees a bound ttp.

gcc/cp/ChangeLog:

PR c++/96531
PR c++/97103
* constraint.cc (map_arguments): Call template_parm_to_arg
in the self-mapping case.
(finish_shorthand_constraint): No need to build a TREE_LIST
before calling template_parm_to_arg.
* pt.c (template_parm_to_arg): Rewrite to handle TEMPLATE_PARM_P
nodes as well as DECL_TEMPLATE_PARM_P nodes, and to make the
overlying TREE_LIST node optional.
(keep_template_parm): Don't record a BOUND_TEMPLATE_TEMPLATE_PARM,
instead record its corresponding TEMPLATE_TEMPLATE_PARM.
(convert_generic_types_to_packs): Don't call
template_parm_to_arg.

gcc/testsuite/ChangeLog:

PR c++/96531
PR c++/97103
* g++.dg/cpp2a/concepts-ttp2.C: New test.
* g++.dg/cpp2a/concepts-variadic1.C: New test.

c++: std::is_constant_evaluated inside constraint [PR97051]

According to [expr.const]/14, the result of substitution into an atomic
constraint is manifestly constant-evaluated; this patch adjusts the call
to maybe_constant_value in satisfy_atom to that effect.

gcc/cp/ChangeLog:

PR c++/97051
* constraint.cc (satisfy_atom): Pass true as the
manifestly_const_eval argument to maybe_constant_value.

gcc/testsuite/ChangeLog:

PR c++/97051
* g++.dg/cpp2a/is-constant-evaluated11.C: New test.

Increase rtx cost of sse_to_integer in skylake_cost.

As https://gcc.gnu.org/pipermail/gcc-patches/2019-August/528839.html
indicates, movement between SSE and gpr should be much expensive than
movement inside gpr(which is 2 as default).

gcc/ChangeLog

PR target/96861
* config/i386/x86-tune-costs.h (skylake_cost): increase rtx
cost of sse_to_integer from 2 to 6.

gcc/testsuite

* gcc.target/i386/pr95021-3.c: Add -mtune=generic.