gcc.git
4 years agolibstdc++: Add comparison operators for string and regex types
Jonathan Wakely [Fri, 17 Apr 2020 15:24:49 +0000 (16:24 +0100)]
libstdc++: Add comparison operators for string and regex types

Some more C++20 changes from P1614R2, "The Mothership has Landed".

This adds three-way comparison support to std::char_traits,
std::basic_string, std::basic_string_view, and std::sub_match.

* include/bits/basic_string.h (basic_string): Define operator<=> and
remove redundant comparison operators for C++20.
* include/bits/char_traits.h (__gnu_cxx::char_traits, char_traits):
Add comparison_category members.
(__detail::__char_traits_cmp_cat): New helper to get comparison
category from char traits class.
* include/bits/regex.h (regex_traits::_RegexMask::operator!=): Do not
define for C++20.
(sub_match): Define operator<=> and remove redundant comparison
operators for C++20.
(match_results): Remove redundant operator!= for C++20.
* include/std/string_view (basic_string_view): Define operator<=> and
remove redundant comparison operators for C++20.
* testsuite/21_strings/basic_string/operators/char/cmp_c++20.cc: New
test.
* testsuite/21_strings/basic_string/operators/wchar_t/cmp_c++20.cc:
New test.
* testsuite/21_strings/basic_string_view/operations/copy/char/
constexpr.cc: Initialize variable.
* testsuite/21_strings/basic_string_view/operations/copy/wchar_t/
constexpr.cc: Likewise.
* testsuite/21_strings/basic_string_view/operators/char/2.cc: Add
dg-do directive and remove comments showing incorrect signatures.
* testsuite/21_strings/basic_string_view/operators/wchar_t/2.cc:
Likewise.
* testsuite/21_strings/basic_string_view/operators/char/cmp_c++20.cc:
New test.
* testsuite/21_strings/basic_string_view/operators/wchar_t/cmp_c++20.cc:
New test.
* testsuite/28_regex/sub_match/compare_c++20.cc: New test.

4 years agoaarch64: Tweak SVE load/store costs
Richard Sandiford [Tue, 14 Apr 2020 20:04:03 +0000 (21:04 +0100)]
aarch64: Tweak SVE load/store costs

We were seeing performance regressions on 256-bit SVE with code like:

  for (int i = 0; i < count; ++i)
  #pragma GCC unroll 128
    for (int j = 0; j < 128; ++j)
      *dst++ = 1;

(derived from lmbench).

For 128-bit SVE, it's clearly better to use Advanced SIMD STPs here,
since they can store 256 bits at a time.  We already do this for
-msve-vector-bits=128 because in that case Advanced SIMD comes first
in autovectorize_vector_modes.

If we handled full-loop predication well for this kind of loop,
the choice between Advanced SIMD and 256-bit SVE would be mostly
a wash, since both of them could store 256 bits at a time.  However,
SVE would still have the extra prologue overhead of setting up the
predicate, so Advanced SIMD would still be the natural choice.

As things stand though, we don't handle full-loop predication well
for this kind of loop, so the 256-bit SVE code is significantly worse.
Something to fix for GCC 11 (hopefully).  However, even though we
account for the overhead of predication in the cost model, the SVE
version (wrongly) appeared to need half the number of stores.
That was enough to drown out the predication overhead and meant
that we'd pick the SVE code over the Advanced SIMD code.

512-bit SVE has a clear advantage over Advanced SIMD, so we should
continue using SVE there.

This patch tries to account for this in the cost model.  It's a bit
of a compromise; see the comment in the patch for more details.

2020-04-17  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.c (aarch64_advsimd_ldp_stp_p): New function.
(aarch64_sve_adjust_stmt_cost): Add a vectype parameter.  Double the
cost of load and store insns if one loop iteration has enough scalar
elements to use an Advanced SIMD LDP or STP.
(aarch64_add_stmt_cost): Update call accordingly.

gcc/testsuite/
* gcc.target/aarch64/sve/cost_model_2.c: New test.
* gcc.target/aarch64/sve/cost_model_3.c: Likewise.
* gcc.target/aarch64/sve/cost_model_4.c: Likewise.
* gcc.target/aarch64/sve/cost_model_5.c: Likewise.
* gcc.target/aarch64/sve/cost_model_6.c: Likewise.
* gcc.target/aarch64/sve/cost_model_7.c: Likewise.

4 years agoc, c++: Fix two redundantAssignment warnings [PR94629]
Jakub Jelinek [Fri, 17 Apr 2020 14:59:57 +0000 (16:59 +0200)]
c, c++: Fix two redundantAssignment warnings [PR94629]

This change fixes two obvious redundant assignments reported by cppcheck:
trunk.git/gcc/c/c-parser.c:16969:2: style: Variable 'data.clauses' is reassigned a value before the old one has been used. [redundantAssignment]
trunk.git/gcc/cp/call.c:5116:9: style: Variable 'arg2' is reassigned a value before the old one has been used. [redundantAssignment]

2020-04-17  Jakub Jelinek  <jakub@redhat.com>

PR other/94629
* c-parser.c (c_parser_oacc_routine): Remove redundant assignment
to data.clauses.

* call.c (build_conditional_expr_1): Remove redundant assignment to
arg2.

4 years agoi386: Fix up *testqi_ext_3 define_insn_and_split [PR94567]
Jakub Jelinek [Fri, 17 Apr 2020 14:56:12 +0000 (16:56 +0200)]
i386: Fix up *testqi_ext_3 define_insn_and_split [PR94567]

As the testcase shows, there are unfortunately more problematic cases
in *testqi_ext_3 if the mode is not CCZmode, because the sign flag might
not behave the same between the insn with zero_extract and what we split it
into.

The previous fix to the insn condition was because *testdi_1 for mask with
upper 32-bits clear and bit 31 set is implemented using SImode test and thus
SF is set depending on that bit 31 rather than on always cleared.

But we can have other cases.  On the zero_extract (which has <MODE>mode),
we can have either the pos + len == precision of <MODE>mode, or
pos + len < precision of <MODE>mode cases.  The former one copies the most
significant bit into SF, the latter will have SF always cleared.

For the former case, either it is a zero_extract from a larger mode, but
then when we perform test in that larger mode, SF will be always clear and
thus mismatch from the zero_extract case (so we need to enforce CCZmode),
or it will be a zero_extract from same mode with pos 0 and len equal to
mode precision, such zero_extracts should have been really simplified
into their first operand.

For the latter case, when SF is always clear on the define_insn with
zero_extract, we need to split into something that doesn't sometimes set
SF, i.e. it has to be a test with mask that doesn't have the most
significant bit set.  In some cases it can be achieved through using test
in a wider mode (e.g. in the testcase, there is
(zero_extract:SI (reg:HI) (const_int 13) (const_int 3))
which will always set SF to 0, but we split it into
(and:HI (reg:HI) (const_int -8))
which will copy the MSB of (reg:HI) into SF, but we can do:
(and:SI (subreg:SI (reg:HI) 0) (const_int 0xfff8))
which will keep SF always cleared), but there are various cases where we
can't (when already using DImode, or when SImode and we'd turned it into
the problematic *testdi_1 implemented using SImode test, or when
the val operand is a MEM (we don't want to read from memory more than
the user originally wanted), paradoxical subreg of MEM could be problematic
too if we through the narrowing end up with a MEM).

So, the patch attempts to require CCZmode (and not CCNOmode) if it can't
really ensure the SF will have same meaning between the define_insn and what
we split it into, and if we decide we allow CCNOmode, it needs to avoid
performing narrowing and/or widen if pos + len would indicate we'd have MSB
set in the mask.

2020-04-17  Jakub Jelinek  <jakub@redhat.com>
    Jeff Law  <law@redhat.com>

PR target/94567
* config/i386/i386.md (*testqi_ext_3): Use CCZmode rather than
CCNOmode in ix86_match_ccmode if len is equal to <MODE>mode precision,
or pos + len >= 32, or pos + len is equal to operands[2] precision
and operands[2] is not a register operand.  During splitting perform
SImode AND if operands[0] doesn't have CCZmode and pos + len is
equal to mode precision.

* gcc.c-torture/execute/pr94567.c: New test.

Co-Authored-By: Jeff Law <law@redhat.com>
4 years agoAvoid testcase warning on arm ABI [pr94608]
Nathan Sidwell [Fri, 17 Apr 2020 13:07:33 +0000 (06:07 -0700)]
Avoid testcase warning on arm ABI [pr94608]

PR c++/94608
* g++.dg/lto/pr83720_0.C: Add fn def to avoid warning on arm ABI.

4 years agoInitialize file_data->lto_section_header before lto_mode_identity_table call.
Martin Liska [Fri, 17 Apr 2020 08:57:03 +0000 (10:57 +0200)]
Initialize file_data->lto_section_header before lto_mode_identity_table call.

PR lto/94612
* lto-common.c: Initialize file_data->lto_section_header
before lto_mode_identity_table call.  It is needed because
it decompresses a LTO section.

4 years agofix PVS studio reported bugs
Richard Biener [Fri, 17 Apr 2020 07:19:32 +0000 (09:19 +0200)]
fix PVS studio reported bugs

2020-04-17  Richard Biener  <rguenther@suse.de>

PR other/94629
* cgraphclones.c (cgraph_node::create_clone): Remove duplicate
initialization.
* dwarf2out.c (dw_val_equal_p): Fix pasto in
dw_val_class_vms_delta comparison.
* optabs.c (expand_binop_directly): Fix pasto in commutation
check.
* tree-ssa-sccvn.c (vn_reference_lookup_pieces): Fix pasto in
initialization.

4 years agoFix -fcompare-debug issue in delete_insn_and_edges [PR94618]
Jakub Jelinek [Fri, 17 Apr 2020 08:33:27 +0000 (10:33 +0200)]
Fix -fcompare-debug issue in delete_insn_and_edges [PR94618]

delete_insn_and_edges calls purge_dead_edges whenever deleting the last insn
in a bb, whatever it is.  If it called it only for mandatory last insns
in the basic block (that may not be followed by DEBUG_INSNs, dunno if that
is control_flow_insn_p or something more complex), that wouldn't be a
problem, but as it calls it on any last insn and can actually do something
in the bb, if such an insn is followed by one more more DEBUG_INSNs and
nothing else in the same bb, we don't call purge_dead_edges with -g and do
call it with -g0.

On the testcase, there are two reg-to-reg moves with REG_EH_REGION notes
(previously memory accesses but simplified and yet not optimized), and the
second is followed by DEBUG_INSNs; the second move is delete_insn_and_edges
and after removing it, for -g0 purge_dead_edges removes the REG_EH_REGION
from the now last insn in the bb (the first reg-to-reg move), while
for -g it isn't called and things diverge from that quickly on.

Fixed by calling purdge_dead_edges even if we remove the last real insn
followed only by DEBUG_INSNs in the same bb.

2020-04-17  Jakub Jelinek  <jakub@redhat.com>

PR rtl-optimization/94618
* cfgrtl.c (delete_insn_and_edges): Set purge not just when
insn is the BB_END of its block, but also when it is only followed
by DEBUG_INSNs in its block.

* g++.dg/opt/pr94618.C: New test.

4 years agoinliner: Don't ICE on NULL TYPE_DOMAIN [PR94621]
Jakub Jelinek [Fri, 17 Apr 2020 07:07:49 +0000 (09:07 +0200)]
inliner: Don't ICE on NULL TYPE_DOMAIN [PR94621]

When I've added the VLA tweak for OpenMP to avoid error_mark_nodes in the IL in
type, I forgot that TYPE_DOMAIN could be NULL.  Furthermore, as an optimization,
this patch checks the hopefully cheapest condition that is very likely false
most of the time (enabled only during OpenMP handling) first.

2020-04-17  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/94621
* tree-inline.c (remap_type_1): Don't dereference NULL TYPE_DOMAIN.
Move id->adjust_array_error_bounds check first in the condition.

* gcc.c-torture/compile/pr94621.c: New test.

4 years agoDo not use HAVE_DOS_BASED_FILE_SYSTEM for Cygwin.
Martin Liska [Fri, 17 Apr 2020 07:22:51 +0000 (09:22 +0200)]
Do not use HAVE_DOS_BASED_FILE_SYSTEM for Cygwin.

PR gcov-profile/94570
* ltmain.sh: Do not define HAVE_DOS_BASED_FILE_SYSTEM
for CYGWIN.

PR gcov-profile/94570
* coverage.c (coverage_init): Use separator properly.

PR gcov-profile/94570
* filenames.h (defined): Do not define HAVE_DOS_BASED_FILE_SYSTEM
for CYGWIN.

Co-Authored-By: Jonathan Yong <10walls@gmail.com>
4 years agotestsuite: Fix up test language requirements [PR94314]
Jakub Jelinek [Fri, 17 Apr 2020 07:01:08 +0000 (09:01 +0200)]
testsuite: Fix up test language requirements [PR94314]

With c++11 one gets:
Excess errors:
.../testsuite/g++.dg/pr94314-4.C:19:28: error: too many arguments to function 'void operator delete(void*)'
because C++ sized deallocation is a C++14 feature.

2020-04-17  Jakub Jelinek  <jakub@redhat.com>

PR c++/94314
* g++.dg/pr94314-4.C: Require c++14 rather than c++11.

4 years agoDo not modify tab options in vimrc for .py files.
Martin Liska [Fri, 17 Apr 2020 06:55:55 +0000 (08:55 +0200)]
Do not modify tab options in vimrc for .py files.

* vimrc: We do not want to modify tab options
for Python files.

4 years agors6000: Fix ICE in decompose_normal_address. [PR93974]
Peter Bergner [Fri, 17 Apr 2020 04:26:41 +0000 (23:26 -0500)]
rs6000: Fix ICE in decompose_normal_address. [PR93974]

Fix an ICE in decompose_normal_address(), which cannot handle Altivec AND:
addresses, by disallowing them via implementing the target hook
rs6000_cannot_substitute_mem_equiv_p.

gcc/
PR rtl-optimization/93974
* config/rs6000/rs6000.c (TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P): Define.
(rs6000_cannot_substitute_mem_equiv_p): New function.

gcc/testsuite/
PR rtl-optimization/93974
* g++.dg/pr93974.C: New test.

4 years agoDaily bump.
GCC Administrator [Fri, 17 Apr 2020 00:16:20 +0000 (00:16 +0000)]
Daily bump.

4 years agocoroutines: Back out mandate for tail-calls at O < 2 [PR94359]
Iain Sandoe [Thu, 16 Apr 2020 13:17:51 +0000 (14:17 +0100)]
coroutines: Back out mandate for tail-calls at O < 2 [PR94359]

For symmetric transfers to work with C++20 coroutines, it is
currently necessary to tail call the callee coroutine from resume
method of the caller coroutine.  However there are several targets
which don't support an indirect tail call to an arbitrary callee.

Unfortunately, the target 'function_ok_for_sibcall' is not usable
from the front end in all cases.  While it is possible to add a new
hook to cover this circumstance, it is too late in the release
cycle to be sure of getting the setting correct for all targets.

So, this patch backs out the use of function_ok_for_sibcall () and
the mandate of CALL_EXPR_MUST_TAIL_CALL from the symmetric
transfer.

Targets that can make indirect tail calls to arbitrary callees will
still be able to make use of the symmetric transfer (without risking
overrunning the stack) for optimization levels >= 2.

The draft standard does not mandate unlimited symmetric transfers,
so removing this is a QOI issue (albeit an important one) rather
than a correctness one.

The test is moved and adjusted so that it can be opted into by any
target that supports the necessary tailcall.

gcc/cp/ChangeLog:

2020-04-16  Iain Sandoe  <iain@sandoe.co.uk>

PR c++/94359
* coroutines.cc (build_actor_fn): Back out use of
targetm.function_ok_for_sibcall.  Do not mark the resume
call as CALL_EXPR_MUST_TAIL_CALL.

gcc/testsuite/ChangeLog:

2020-04-16  Iain Sandoe  <iain@sandoe.co.uk>

PR c++/94359
* g++.dg/coroutines/torture/symmetric-transfer-00-basic.C: Move..
* g++.dg/coroutines/symmetric-transfer-00-basic.C: ..here and
adjust to run at O2 for targets supporting the necessary tail
call.

4 years agotestsuite: Update pr94426-2 test to handle more targets [PR94426]
Iain Sandoe [Thu, 16 Apr 2020 20:13:51 +0000 (21:13 +0100)]
testsuite: Update pr94426-2 test to handle more targets [PR94426]

The function (_Z4InitIN3VARIZ1qvEUlvE_EUlvE_EEbT_) that initializes
VAR has changed siganture.  It was was weak / comdat [Linux] or
weak / global [Darwin] and now is text section local.

So: test that the symbol exists, is spelled the way intended by the
ABI mangling and is not weak (or global on Darwin).

gcc/testsuite/ChangeLog:

2020-04-16  Iain Sandoe  <iain@sandoe.co.uk>

PR c++/94426
* g++.dg/cpp0x/lambda/pr94426-2.C: Adjust scan-asms to test
for the change on more platforms.

4 years agoipa: Make call redirection detect already adjusted calls (PR 93621)
Martin Jambor [Thu, 16 Apr 2020 17:21:02 +0000 (19:21 +0200)]
ipa: Make call redirection detect already adjusted calls (PR 93621)

PR 93621 testcase makes redirect_call_stmt_to_callee wrongly assume
that a call statement needs redirecting but then rightly fails an
assert ensuring the call statement parameters have not already been
adjusted because they were already created adjusted as part of thunk
expansion.

The test fails because the decl in the call call statement is
different than the decl of the callee, because the latter was created
in save_inline_function_body.  This patch adds a way to link these two
and detect the situation in redirect_call_stmt_to_callee.

2020-04-16  Martin Jambor  <mjambor@suse.cz>

PR ipa/93621
* ipa-inline.h (ipa_saved_clone_sources): Declare.
* ipa-inline-transform.c (ipa_saved_clone_sources): New variable.
(save_inline_function_body): Link the new body holder with the
previous one.
* cgraph.c: Include ipa-inline.h.
(cgraph_edge::redirect_call_stmt_to_callee): Try to find the decl from
the statement in ipa_saved_clone_sources.
* cgraphunit.c: Include ipa-inline.h.
(expand_all_functions): Free ipa_saved_clone_sources.

testsuite/
* g++.dg/ipa/pr93621.C: New test.

4 years agoaarch64: Fix mismatched SVE predicate modes [PR94606]
Richard Sandiford [Wed, 15 Apr 2020 12:52:20 +0000 (13:52 +0100)]
aarch64: Fix mismatched SVE predicate modes [PR94606]

For this testcase we ended up generating the invalid rtl:

(insn 10 9 11 2 (set (reg:VNx16BI 105)
        (and:VNx16BI (xor:VNx16BI (reg:VNx8BI 103)
                (reg:VNx16BI 104))
            (reg:VNx16BI 104))) "/tmp/bar.c":9:12 -1
     (nil))

Fixed by taking the VNx16BI lowpart.  It's safe to do that here because
the gp (r104) masks out the extra odd-indexed bits.

2020-04-16  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR target/94606
* config/aarch64/aarch64.c (aarch64_expand_sve_const_pred_eor): Take
the VNx16BI lowpart of the recursively-generated constant.

gcc/testsuite/
PR target/94606
* gcc.dg/vect/pr94606.c: New test.

4 years agoList valid pairs for new and delete operators.
Martin Liska [Thu, 16 Apr 2020 13:39:22 +0000 (15:39 +0200)]
List valid pairs for new and delete operators.

PR c++/94314
* cgraphclones.c (set_new_clone_decl_and_node_flags): Drop
DECL_IS_REPLACEABLE_OPERATOR during cloning.
* tree-ssa-dce.c (valid_new_delete_pair_p): New function.
(propagate_necessity): Check operator names.

PR c++/94314
* g++.dg/pr94314.C: Do not use dg-additional-options
and remove not needed stdio.h include.
* g++.dg/pr94314-2.C: Likewise.
* g++.dg/pr94314-3.C: Likewise.
* g++.dg/pr94314-4.C: New test.

Co-Authored-By: Jakub Jelinek <jakub@redhat.com>
4 years agoc++: Error recovery with erroneous DECL_INITIAL [PR94475]
Patrick Palka [Wed, 15 Apr 2020 16:47:41 +0000 (12:47 -0400)]
c++: Error recovery with erroneous DECL_INITIAL [PR94475]

Here we're ICE'ing in do_narrow during error-recovery, because ocp_convert
returns error_mark_node after it attempts to reduce a const decl to its
erroneous DECL_INITIAL via scalar_constant_value, and we later pass this
error_mark_node to fold_build2 which isn't prepared to handle error_mark_nodes.

We could fix this ICE in do_narrow by checking if ocp_convert returns
error_mark_node, but for the sake of consistency and for better error recovery
it seems it'd be preferable if ocp_convert didn't care that a const decl's
initializer is erroneous and would instead proceed as if the decl was not const,
which is the approach that this patch takes.

gcc/cp/ChangeLog:

PR c++/94475
* cvt.c (ocp_convert): If the result of scalar_constant_value is
erroneous, ignore it and use the original expression.

gcc/testsuite/ChangeLog:

PR c++/94475
* g++.dg/conversion/err-recover2.C: New test.
* g++.dg/diagnostic/pr84138.C: Remove now-bogus warning.
* g++.dg/warn/Wsign-compare-8.C: Remove now-bogus warning.

4 years agoMake it possible to have different instrumented and feedback builds without copying...
Jan Hubicka [Thu, 16 Apr 2020 12:25:51 +0000 (14:25 +0200)]
Make it possible to have different instrumented and feedback builds without copying gcda files around [pr93401]

Hi,
in GCC 8 we changed -fprofile-generate=<path> to use mangled absolute paths in
the <path> directory. This was necessary to avoid clashes of files when gcc is
executed from different directories to build different sources of same
filename.

However this made it difficult to build projects on setups where instrumented
build is done in one directory, feedback build in different and possibly
training happens in yet another directory structure.  This happens i.e. for
Firefox builds for month or two.

This patch adds -fprofile-prefix-path that can be used to inform gcc where the
root of build directory is and strip it form the gcda filenames.
This is similar to exisitng debug-prefix-map but without the map feature since
it seems useless for profile data.

We spent quite some time with Maritn Liska discussing options and found no
better solution.  I was looking how this work on LLVM and they produce single
profdata file which is then transformed into kind of simple database by
llvmprofdata tool.  This database keys functions by filename and symbol name.
If you arrane two files with same name define static variable with same symbol
name this gets messedup and result in wrong info. So I think this is not very
good solution and preffer the extra option.

Bootstrapped/regtested x86_64-linux. I plan to commit it later today if there
are no complains.

I suppose our manual could have some central section on profile feedback
explaining the whole setup at one place.

Honza

PR gcov-profile/93401
* common.opt (profile-prefix-path): New option.
* coverae.c: Include diagnostics.h.
(coverage_init): Strip profile prefix path.
* doc/invoke.texi (-fprofile-prefix-path): Document.

4 years agoearly-remat: Handle sets of multiple candidate regs [PR94605]
Richard Sandiford [Wed, 15 Apr 2020 10:38:24 +0000 (11:38 +0100)]
early-remat: Handle sets of multiple candidate regs [PR94605]

early-remat.c:process_block wasn't handling insns that set multiple
candidate registers, which led to an assertion failure at the end
of the main loop.

Instructions that set two pseudos aren't rematerialisation candidates in
themselves, but we still need to track them if another instruction that
sets the same register is a rematerialisation candidate.

2020-04-16  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR rtl-optimization/94605
* early-remat.c (early_remat::process_block): Handle insns that
set multiple candidate registers.

gcc/testsuite/
PR rtl-optimization/94605
* gcc.target/aarch64/sve/pr94605.c: New test.

4 years agocleanup graphite results
Richard Biener [Thu, 16 Apr 2020 11:07:51 +0000 (13:07 +0200)]
cleanup graphite results

This removes { dg-final { scan-tree-dump "tiled" "graphite" } } scans
from graphite tests that pass/fail dependent on the ISL version used.
Note all scans match the actually dumped "not tiled" messages with
ISL 0.12 and not the alternative "tiled by <number>".  With ISL
0.22 neither is printed because the tiling infrastructure doesn't
expect the new schedule layout (though looking at a few cases it
looks superior with more dimensions marked as permutable.

Anyway, the scans have nothing to do with interchange and just
add to testsuite noise.

2020-04-16  Richard Biener  <rguenther@suse.de>

* gcc.dg/graphite/interchange-1.c: Remove scan for tiled.
* gcc.dg/graphite/interchange-10.c: Likewise.
* gcc.dg/graphite/interchange-11.c: Likewise.
* gcc.dg/graphite/interchange-3.c: Likewise.
* gcc.dg/graphite/interchange-4.c: Likewise.
* gcc.dg/graphite/interchange-7.c: Likewise.
* gcc.dg/graphite/interchange-9.c: Likewise.
* gcc.dg/graphite/uns-interchange-9.c: Likewise.
* gfortran.dg/graphite/interchange-3.f90: Likewise.

4 years agomiddle-end/94614 - avoid multiword moves to nothing
Richard Biener [Thu, 16 Apr 2020 07:37:44 +0000 (09:37 +0200)]
middle-end/94614 - avoid multiword moves to nothing

This adjusts emit_move_multi_word to handle moves into paradoxical
subregs parts that are not there and adjusts lower-subregs
CLOBBER resolving to deal with those as well.

2020-04-16  Richard Biener  <rguenther@suse.de>

PR middle-end/94614
* expr.c (emit_move_multi_word): Do not generate code when
the destination part is undefined_operand_subword_p.
* lower-subreg.c (resolve_clobber): Look through a paradoxica
subreg.

4 years agointl: Unbreak intl build with bison 3 when no regeneration is needed [PR92008]
Jakub Jelinek [Thu, 16 Apr 2020 09:55:00 +0000 (11:55 +0200)]
intl: Unbreak intl build with bison 3 when no regeneration is needed [PR92008]

As Iain reported, my change broke the case when one has bison >= 3,
but make decides there is no reason to regenerate plural.c, unfortunately
that seems to be a scenario I haven't tested.  The problem is that
the pregenerated plural.c has been generated with bison 1.35, but when
config.h says HAVE_BISON3, the code assumes it is the bison3 variant.
What used to work fine is when one has bison >= 3 and plural.c has been
regenerated (e.g. do touch intl/plural.y and it will work), or when
one doesn't have any bison (then nothing is regenerated, but HAVE_BISON3
isn't defined either), or when one has bison < 3 and doesn't need to
regenerate, or when one has bison < 3 and it is regenerated.

The following patch fixes this, by killing the HAVE_BISON3 macro from
config.h, and instead remembering the fact whether plural.c has been created
with bison < 3 or bison >= 3 in a separate new plural-config.h header.
The way this works:
- user doesn't have bison
- user has bison >= 3, but intl/{plural-config.h,plural.c} aren't older than intl/plural.y
- user has bison < 3, but intl/{plural-config.h,plural.c} aren't older than intl/plural.y
pregenerated !USE_BISON3 plural.c and plural-config.h from source
dir is used, nothing in the objdir
- user has bison >= 3 and intl/plural.y is newer
Makefile generates plural.c and USE_BISON3 plural-config.h in the
objdir, which is then used in preference to srcdir copies
- user has bison < 3 and intl/plural.y is newer
Makefile generates plural.c and !USE_BISON3 plural-config.h in the
objdir, which is then used in preference to srcdir copies
I have tested all these cases and make all-yes worked in all the cases.
If one uses the unsupported ./configure where srcdir == objdir, I guess
(though haven't tested) that it should still work, just it would be nice
if such people didn't try to check in the plural{.c,-config.h} they have
regenerated.
What doesn't work, but didn't work before either (just tested gcc-9 branch
too) is when one doesn't have bison and plural.y is newer than plural.c.
Don't do that ;)

2020-04-16  Jakub Jelinek  <jakub@redhat.com>

PR bootstrap/92008
intl/
* configure.ac: Remove HAVE_BISON3 AC_DEFINE.
* Makefile.in (HEADERS): Add plural-config.h.
(.y.c): Also create plural-config.h.
(dcigettext.o loadmsgcat.o plural.o plural-exp.o): Also depend
on plural-config.h.
(plural-config.h): Depend on plural.c.
* plural-exp.h: Include plural-config.h.  Use USE_BISON3 instead
of HAVE_BISON3.
* plural.y: Use USE_BISON3 instead of HAVE_BISON3.
* configure: Regenerated.
* plural.c: Regenerated.
* config.h.in: Regenerated.
* plural-config.h: Generated.
contrib/
* gcc_update: Add intl/plural.y dependency for intl/plural-config.h.

4 years agosra: Fix access verification (PR 94598)
Martin Jambor [Thu, 16 Apr 2020 09:04:41 +0000 (11:04 +0200)]
sra: Fix access verification (PR 94598)

get_ref_base_and_extent recognizes ARRAY_REFs with variable index but
into arrays of length one as constant offset accesses.  However,
max_size in such cases is extended to span the whole element.  This
confuses SRA verification when SRA also builds its (total
scalarization) access structures to describe fields under such array -
get_ref_base_and_extent returns different size and max_size for them.

Fixed by not performing the check for total scalarization accesses.
The subsequent check then had to be changed to use size and not
max_size too, which meant it has to be skipped when the access
structure describes a genuine variable array access.

Bootstrapped and tested on x86_64-linux.

2020-04-16  Martin Jambor  <mjambor@suse.cz>

PR tree-optimization/94598
* tree-sra.c (verify_sra_access_forest): Fix verification of total
scalarization accesses under access to one-element arrays.

testsuite/
* gcc.dg/tree-ssa/pr94598.c: New test.

4 years agotestsuite: Move misplaced gcc.c-torture/pr92372.c test [PR92372]
Jakub Jelinek [Thu, 16 Apr 2020 08:37:46 +0000 (10:37 +0200)]
testsuite: Move misplaced gcc.c-torture/pr92372.c test [PR92372]

This test got committed into a spot where nothing actually tests it.
As there is no main, I assume it was meant to be gcc.c-torture/compile/
test and the test PASSes after moving there (both x86_64-linux and
i686-linux).  Though, it passed before the PR92372 fixes too.

2020-04-16  Jakub Jelinek  <jakub@redhat.com>

PR ipa/92372
* gcc.c-torture/pr92372.c: Move ...
* gcc.c-torture/compile/pr92372.c: ... here.

4 years agobootstrap: Fix building with GCC 4.2 [PR89494]
Jakub Jelinek [Thu, 16 Apr 2020 08:15:18 +0000 (10:15 +0200)]
bootstrap: Fix building with GCC 4.2 [PR89494]

GCC 4.2 (but I think not the latest tip of GCC 4.2 branch) has broken value
initialization, see PR33916.  The following patch provides a workaround for
that.  Tested with GCC 4.2 on a reduced testcase I've distilled from the
assign_param_data_one class which has been miscompiled the same.

2020-04-16  Jakub Jelinek  <jakub@redhat.com>

PR bootstrap/89494
* function.c (assign_parm_find_data_types): Add workaround for
BROKEN_VALUE_INITIALIZATION compilers.

4 years agointl: Allow building both with old bison and bison >= 3 [PR92008]
Jakub Jelinek [Thu, 16 Apr 2020 08:12:30 +0000 (10:12 +0200)]
intl: Allow building both with old bison and bison >= 3 [PR92008]

bison 3 apparently made a backwards incompatible change, dropped
YYLEX_PARAM/YYPARSE_PARAM support and instead needs %param or %lex-param
and %parse-param.  Furthermore, there is no easy way to conditionalize
on bison version in the *.y files.
While e.g. glibc bumped bison requirement and just has the bison 3
compatible version, Richi said there are still systems with older bison
where we want to build gcc.

So, this patch instead determines during configure bison version, and
depending on that when building plural.c (if building it at all) tweaks
what is passed over to bison if needed.

Tested with both bison 3 and bison 1.35, in each case with reconfiguring
intl and building with make all-yes (as in my setup intl isn't normally
used).

2020-04-16  Jakub Jelinek  <jakub@redhat.com>

PR bootstrap/92008
* configure.ac: Add check for bison >= 3, AC_DEFINE HAVE_BISON3
and AC_SUBST BISON3_YES and BISON3_NO.
* Makefile.in (.y.c): Prefix $(YACC) invocation with @BISON3_NO@,
add @BISON3_YES@ prefixed rule to adjust the *.y source using sed
and adjust output afterwards.
* plural-exp.h (PLURAL_PARSE): If HAVE_BISON3 is defined, use
struct parse_args * type for arg instead of void *.
* plural.y: Add magic /* BISON3 ... */ comments with bison >= 3
directives.
(YYLEX_PARAM, YYPARSE_PARAM): Don't define if HAVE_BISON3 is defined.
(yylex, yyerror): Adjust prototypes and definitions if HAVE_BISON3
is defined.
* plural.c: Regenerated.
* config.h.in: Regenerated.
* configure: Regenerated.

4 years agopretty-print SSA names
Richard Biener [Wed, 15 Apr 2020 12:50:00 +0000 (14:50 +0200)]
pretty-print SSA names

This adds the SSA name version to the gdb pretty-printing of SSA names.

(gdb) p (tree)$1
$5 = <ssa_name 0x7ffff68435a0 323>

2020-04-16  Richard Biener  <rguenther@suse.de>

* gdbhooks.py (TreePrinter): Print SSA_NAME_VERSION of SSA_NAME
nodes.

4 years agolibstdc++: Fix -Wunused-parameter warning in test
Jonathan Wakely [Thu, 16 Apr 2020 07:44:10 +0000 (08:44 +0100)]
libstdc++: Fix -Wunused-parameter warning in test

* testsuite/20_util/unsynchronized_pool_resource/allocate.cc: Remove
name of unused parameter.

4 years agoc++: Fix pasto in structured binding diagnostics [PR94571]
Jakub Jelinek [Thu, 16 Apr 2020 05:19:57 +0000 (07:19 +0200)]
c++: Fix pasto in structured binding diagnostics [PR94571]

This snippet has been copied from the non-structured binding declaration
parsing later in the function, and while for non-structured bindings
it can be followed by comma or semicolon, structured bindings may be
only followed by semicolon.

Or, do we want to have a different message for the case when there is
a comma (and keep this corrected one only if there is something else)
that would explain better what is the bug (or add a fix-it hint)?
Marek said in the PR that clang++ reports
error: decomposition declaration must be the only declaration in its group

There is another thing Marek noted (though, something for different spot),
that diagnostic for auto x(1), [e,f] = test2; could also use a clearer
wording like the above (or a fix-it hint), but the question is if we should
assume [ after , as a structured binding or if we should do some tentative
parsing first to figure out if it looks like a structured binding.

2020-04-16  Jakub Jelinek  <jakub@redhat.com>

PR c++/94571
* parser.c (cp_parser_simple_declaration): Fix up a pasto in
diagnostics.

* g++.dg/cpp1z/decomp51.C: New test.

4 years agoDaily bump.
GCC Administrator [Thu, 16 Apr 2020 00:16:24 +0000 (00:16 +0000)]
Daily bump.

4 years agoruntime: use 64 bits of hash seed on arm64
eric fang [Thu, 20 Feb 2020 06:57:11 +0000 (06:57 +0000)]
runtime: use 64 bits of hash seed on arm64

This is the same issue as #33960, but on gofrontend.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/220297

4 years agolibstdc++: Add comparison operators to std::filesystem types
Jonathan Wakely [Wed, 15 Apr 2020 20:01:42 +0000 (21:01 +0100)]
libstdc++: Add comparison operators to std::filesystem types

Some more C++20 changes from P1614R2, "The Mothership has Landed".

* include/bits/fs_dir.h (file_status): Define operator== for C++20.
(directory_entry): Define operator<=> and remove redundant comparison
operators for C++20.
* include/bits/fs_fwd.h (space_info): Define operator== for C++20.
* include/bits/fs_path.h (path): Define operator<=> and remove
redundant comparison operators for C++20.
* testsuite/27_io/filesystem/path/compare/compare.cc: Fix comment.
* testsuite/27_io/filesystem/path/compare/lwg2936.cc: Likewise.
* testsuite/27_io/filesystem/path/compare/path.cc: Likewise.
* testsuite/27_io/filesystem/path/compare/strings.cc: Likewise.

4 years agoGo Gcc_backend class: mark prefetch as novops
Ian Lance Taylor [Wed, 15 Apr 2020 19:59:52 +0000 (12:59 -0700)]
Go Gcc_backend class: mark prefetch as novops

PR go/94607
* go-gcc.cc (class Gcc_backend): Define builtin_const,
builtin_noreturn, builtin_novops.
(Gcc_backend::define_builtin): Change const_p and noreturn_p
parameters to a single flags parameter.  Change all callers.
(Gcc_backend::Gcc_backend): Pass novops for prefetch.

4 years agolibstdc++: Add comparison operators to types from Utilities clause
Jonathan Wakely [Wed, 15 Apr 2020 18:47:48 +0000 (19:47 +0100)]
libstdc++: Add comparison operators to types from Utilities clause

Some more C++20 changes from P1614R2, "The Mothership has Landed".

This removes all redundant equality and inequality operators in the
Utilities clause, as they can be synthesized from the remaining equality
operators.

It also removes the single redundant operator in the Localization
clause, because it didn't seem worth doing in a separate commit.

* include/bits/allocator.h (operator!=): Do not define for C++20.
* include/bits/locale_classes.h (operator!=): Likewise.
* include/bits/std_function.h (operator==(nullptr_t, const function&))
(operator!=(const function&, nullptr_t))
(operator!=(nullptr_t, const function&)): Likewise.
* include/ext/bitmap_allocator.h (operator!=): Likewise.
* include/ext/debug_allocator.h (operator!=): Likewise.
* include/ext/extptr_allocator.h (operator!=): Likewise.
* include/ext/malloc_allocator.h (operator!=): Likewise.
* include/ext/mt_allocator.h (operator!=): Likewise.
* include/ext/new_allocator.h (operator!=): Likewise.
* include/ext/pool_allocator.h (operator!=): Likewise.
* include/ext/throw_allocator.h (operator!=): Likewise.
* include/std/bitset (bitset::operator!=): Likewise.
* include/std/memory_resource (operator!=): Likewise.
* include/std/scoped_allocator (operator!=): Likewise.

4 years agolibstdc++: Add spaceship operator to std::type_index
Jonathan Wakely [Wed, 15 Apr 2020 18:47:47 +0000 (19:47 +0100)]
libstdc++: Add spaceship operator to std::type_index

Another C++20 change from P1614R2, "The Mothership has Landed".

* include/std/typeindex (operator<=>): Define for C++20.
* testsuite/20_util/typeindex/comparison_operators_c++20.cc: New test.

4 years agoSimplify recently introduced check to simplify_bound.
Fritz Reese [Wed, 15 Apr 2020 15:44:40 +0000 (11:44 -0400)]
Simplify recently introduced check to simplify_bound.

gcc/fortran/ChangeLog:

2020-04-15  Fritz Reese  <foreese@gcc.gnu.org>
Linus Koenig <link@sig-st.de>

PR fortran/94192
* simplify.c (simplify_bound): Simplify condition for error check.

4 years agoi386: Require OPTION_MASK_ISA_SSE2 for __builtin_ia32_movq128 [PR94603]
Uros Bizjak [Wed, 15 Apr 2020 15:08:07 +0000 (17:08 +0200)]
i386: Require OPTION_MASK_ISA_SSE2 for __builtin_ia32_movq128 [PR94603]

PR target/94603
* config/i386/i386-builtin.def (__builtin_ia32_movq128):
Require OPTION_MASK_ISA_SSE2.

testsuite/ChangeLog:

PR target/94603
* gcc.target/i386/pr94603.c: New test.

4 years agoselftest: Work around GCC 4.2 PR33916 bug by optimizing the ctor [PR89494]
Gustavo Romero [Wed, 15 Apr 2020 13:14:45 +0000 (15:14 +0200)]
selftest: Work around GCC 4.2 PR33916 bug by optimizing the ctor [PR89494]

GCC 4.2 due to PR33916 miscompiles temp_dump_context ctor, because it doesn't
zero initialize the whole dump_context temporary on which it runs the static
get method and during destruction of the temporary an uninitialized pointer
is deleted.

More recent GCC versions properly zero initialize it and ideally optimize away
the construction/destruction of the temporary, as it isn't used for anything,
but there is no reason to create the temporary, static member functions can
be called without an associated object.

2020-04-15  Gustavo Romero  <gromero@linux.ibm.com>

PR bootstrap/89494
* dumpfile.c (selftest::temp_dump_context::temp_dump_context):
Don't construct a dump_context temporary to call static method.

4 years agoaarch64: Fix valid_src_p for use of uninitialized value
Andrea Corallo [Thu, 9 Apr 2020 14:34:50 +0000 (15:34 +0100)]
aarch64: Fix valid_src_p for use of uninitialized value

2020-04-15  Andrea Corallo  <andrea.corallo@arm.com>

* config/aarch64/falkor-tag-collision-avoidance.c
(valid_src_p): Check for aarch64_address_info type before
accessing base field.

4 years agoArm: MVE: Add mve vec_duplicate pattern
Andre Vieira [Wed, 15 Apr 2020 11:53:46 +0000 (12:53 +0100)]
Arm: MVE: Add mve vec_duplicate pattern

This patch fixes an ICE we were seeing due to a missing vec_duplicate pattern.

gcc/ChangeLog:
2020-04-15  Andre Vieira  <andre.simoesdiasvieira@arm.com>

* config/arm/mve.md (mve_vec_duplicate<mode>): New pattern.
(V_sz_elem2): Remove unused mode attribute.

gcc/testsuite/ChangeLog:
2020-04-15  Andre Vieira <andre.simoesdiasvieira@arm.com>
            Srinath Parvathaneni <srinath.parvathaneni@arm.com>

* gcc.target/arm/mve/intrinsics/mve_vec_duplicate.c: New test.

4 years ago[Arm] Disallow arm_movdi when targetting MVE
Matthew Malcomson [Wed, 15 Apr 2020 11:24:39 +0000 (12:24 +0100)]
[Arm] Disallow arm_movdi when targetting MVE

Without disabling this, the pattern can try and move DImode values
between floating point registers and general registers.
The constraints on this pattern can't handle that, and reload goes into
an infinite loop.

This was the cause of a testsuite failure in cde_v_1_mve.c, which is now
gone.

A DImode move for MVE now uses the `movdi_vfp` pattern, which is the same
pattern used for such a move when MVE is not available but the target has
TARGET_HARD_FLOAT.

Testing done:
    Bootstrapped and regtested on arm-none-linux-gnueabihf
    regtested on arm-none-eabi

gcc/ChangeLog:

2020-04-15  Matthew Malcomson  <matthew.malcomson@arm.com>

* config/arm/arm.md (arm_movdi): Disallow for MVE.

4 years agoaarch64: Fix bootstrap with old binutils [PR93053]
Jakub Jelinek [Wed, 15 Apr 2020 09:01:19 +0000 (11:01 +0200)]
aarch64: Fix bootstrap with old binutils [PR93053]

As reported in the PR, GCC 10 (and also 9.3.1 but not 9.3.0) fails to build
when using older binutils which lack LSE support, because those instructions
are used in libgcc.
Thanks to Kyrylo's hint, the following patches (hopefully) allow it to build
even with older binutils by using .inst directive if LSE support isn't
available in the assembler.

2020-04-15  Jakub Jelinek  <jakub@redhat.com>

PR target/93053
* configure.ac (LIBGCC_CHECK_AS_LSE): Add HAVE_AS_LSE checking.
* config/aarch64/lse.S: Include auto-target.h, if HAVE_AS_LSE
is not defined, use just .arch armv8-a.
(B, M, N, OPN): Define.
(COMMENT): New .macro.
(CAS, CASP, SWP, LDOP): Use .inst directive if HAVE_AS_LSE is not
defined.  Otherwise, move the operands right after the glue? and
comment out operands where the macros are used.
* configure: Regenerated.
* config.in: Regenerated.

4 years agoopenmp: Reject requires directives not at file or namespace scope [PR94593]
Jakub Jelinek [Wed, 15 Apr 2020 07:59:14 +0000 (09:59 +0200)]
openmp: Reject requires directives not at file or namespace scope [PR94593]

This change started with a bugreport about a typo in one requires testcase
(diagnosed with -Wunknown-pragmas only), but following discussion lead to
noting that we do not diagnose restriction that requires directives in
C/C++ may only appear at file or namespace scope; and several our tests
violated that.

2020-04-15  Jakub Jelinek  <jakub@redhat.com>

PR c/94593
* c-parser.c (c_parser_pragma) <case PRAGMA_OMP_REQUIRES>: Reject
requires directive when not at file scope.

* parser.c (cp_parser_pragma) <case PRAGMA_OMP_REQUIRES>: Reject
requires directive when not at file or namespace scope.

* c-c++-common/gomp/requires-1.c: Fix a typo, requries -> requires.
Move directives to file scope.
(i): Remove.
* c-c++-common/gomp/requires-2.c: Move directives to file scope.
(i, foo): Remove.
* c-c++-common/gomp/requires-4.c: Move directives to file scope.
* c-c++-common/gomp/atomic-19.c: Move requires directive to file scope.
* c-c++-common/gomp/atomic-20.c: Likewise.
* c-c++-common/gomp/atomic-21.c: Likewise.
* c-c++-common/gomp/atomic-22.c: Likewise.
* gcc.dg/gomp/requires-1.c: New test.
* g++.dg/gomp/requires-1.C: New test.
* g++.dg/gomp/requires-2.C: New test.
* g++.dg/gomp/atomic-18.C: Move requires directive to file scope.

4 years agomiddle-end/94539 - void * aliases every other pointer
Richard Biener [Tue, 14 Apr 2020 11:16:25 +0000 (13:16 +0200)]
middle-end/94539 - void * aliases every other pointer

This makes same_type_for_tbaa_p conservative in the same way
get_alias_set is about void * which we allow to alias all other
pointers.

2020-04-15  Richard Biener  <rguenther@suse.de>

PR middle-end/94539
* tree-ssa-alias.c (same_type_for_tbaa): Defer to
alias_sets_conflict_p for pointers.

* gcc.dg/alias-14.c: Make dg-do run.

4 years agoDaily bump.
GCC Administrator [Wed, 15 Apr 2020 00:16:20 +0000 (00:16 +0000)]
Daily bump.

4 years agoxtensa: fix PR target/94584
Max Filippov [Mon, 13 Apr 2020 20:26:04 +0000 (13:26 -0700)]
xtensa: fix PR target/94584

Patterns zero_extendhisi2, zero_extendqisi2 and extendhisi2_internal can
load value from memory, but they don't treat volatile memory correctly.
Add %v1 before load instructions to emit 'memw' instruction when
-mserialize-volatile is in effect.

2020-04-14  Max Filippov  <jcmvbkbc@gmail.com>
gcc/
* config/xtensa/xtensa.md (zero_extendhisi2, zero_extendqisi2)
(extendhisi2_internal): Add %v1 before the load instructions.

gcc/testsuite/
* gcc.target/xtensa/pr94584.c: New test.

4 years agoUpdate gcc sv.po.
Joseph Myers [Tue, 14 Apr 2020 23:51:21 +0000 (23:51 +0000)]
Update gcc sv.po.

* sv.po: Update.

4 years agolibstdc++: Fix constraints on std::compare_three_way
Jonathan Wakely [Tue, 14 Apr 2020 20:58:03 +0000 (21:58 +0100)]
libstdc++: Fix constraints on std::compare_three_way

My "simplification" of std::compare_three_way's constraints in commit
f214ffb336d582a66149068a2a96b7fcf395b5de was incorrect, because
std::three_way_comparable_with imposes additional restrictions beyond
the <=> expression being valid.

* libsupc++/compare (compare_three_way): Fix constraint so that
BUILTIN-PTR-THREE-WAY does not require three_way_comparable_with.
* testsuite/18_support/comparisons/object/builtin-ptr-three-way.cc:
New test.

4 years agolibstdc++: Add comparison operators to std::shared_ptr (PR 94562)
Jonathan Wakely [Tue, 14 Apr 2020 20:54:55 +0000 (21:54 +0100)]
libstdc++: Add comparison operators to std::shared_ptr (PR 94562)

This also implements the proposed resolution to LWG issue 3247, so that
the ill-formed <=> expression with nullptr is not used.

PR libstdc++/94562
* include/bits/shared_ptr.h (operator<=>): Define for C++20.
* include/bits/shared_ptr_base.h (operator<=>): Likewise.
* include/bits/unique_ptr.h (operator<=>): Add inline specifier.
* testsuite/20_util/shared_ptr/comparison/cmp_c++20.cc: New test.
* testsuite/20_util/shared_ptr/comparison/less.cc: Do not expect
std::less<A*> to be used when comparing std::shared_ptr<A> objects in
C++20.

4 years agocoroutines: Fix compile error with symmetric transfers [PR94359]
Iain Sandoe [Tue, 14 Apr 2020 19:37:12 +0000 (20:37 +0100)]
coroutines: Fix compile error with symmetric transfers [PR94359]

For symmetric transfers to work with C++20 coroutines, it is
currently necessary to tail call the callee coroutine from resume
method of the caller coroutine. The current codegen marks these
resume calls as "MUST_TAIL_CALL" to indicate that the tail call is
required for correctness.

Unfortunately, several targets have ABI constraints that prevent
an indirect tail-call, which results in the PRs compile error.

The change here tests the target sibcall hook for the resume
expression and only marks it as requiring a tail call if that's
supported.

This doesn't fix the underlying problem; that really a solution is
needed to allow the tail-calls (or equivalent) to take place - but
that will be deferred until next stage 1.

gcc/cp/ChangeLog:

2020-04-14  Iain Sandoe  <iain@sandoe.co.uk>

PR c++/94359
* coroutines.cc (build_actor_fn): Check that the target can
support the resume tailcall before mandating it.

gcc/testsuite/ChangeLog:

2020-04-14  Iain Sandoe  <iain@sandoe.co.uk>

PR c++/94359
* g++.dg/coroutines/torture/symmetric-transfer-00-basic.C:
Expect a run fail for targets without arbitrary indirect
tail-calls.

4 years agolibstdc++: Fix order of entries in ChangeLog
Jonathan Wakely [Tue, 14 Apr 2020 19:42:48 +0000 (20:42 +0100)]
libstdc++: Fix order of entries in ChangeLog

4 years agoFix for PR/94542, do not make TLS refs PC-relative
Aaron Sawdey [Tue, 14 Apr 2020 19:38:47 +0000 (14:38 -0500)]
Fix for PR/94542, do not make TLS refs PC-relative

For rs6000 target, it is not valid to make PC-relative
references to TLS symbols. So addr_to_insn_form() needs
to check if things are TLS before returning PC-rel forms.

2020-04-14  Aaron Sawdey  <acsawdey@linux.ibm.com>

PR target/94542
* config/rs6000/rs6000.c (address_to_insn_form): Do not attempt to
use PC-relative addressing for TLS references.

4 years agoc++: Add testcase for PR c++/93207
Patrick Palka [Tue, 14 Apr 2020 19:02:22 +0000 (15:02 -0400)]
c++: Add testcase for PR c++/93207

gcc/testsuite/ChangeLog:

PR c++/93207
* g++.dg/concepts/variadic5.C: New test.

4 years agoc++: "'decltype_type' not supported" in diagnostic [PR85278]
Patrick Palka [Tue, 14 Apr 2020 18:22:31 +0000 (14:22 -0400)]
c++: "'decltype_type' not supported" in diagnostic [PR85278]

This fixes a garbled concepts diagnostic by moving the handling of DECLTYPE_TYPE
from pp_cxx_type_specifier_seq to cxx_pretty_printer::simple_type_specifier, a
move which also seems to be more consistent with the language grammar.

This patch also fixes pretty printing of rvalue reference types via
cxx_pretty_printer::type_id, which eventually calls pp_c_pointer which currently
doesn't distinguish between lvalue and rvalue references.

gcc/c-family/ChangeLog:

PR c++/85278
* c-pretty-print.c (pp_c_pointer) <case REFERENCE_TYPE>: Print a double
ampersand if it's an rvalue reference type.

gcc/cp/ChangeLog:

PR c++/85278
* cxx-pretty-print.c (cxx_pretty_printer:simple_type_specifier)
<case DECLTYPE_TYPE>: Handle DECLTYPE_TYPE here instead of ...
(pp_cxx_type_specifier_seq) <case DECLTYPE_TYPE>: ... here.
(cxx_pretty_printer::direct_abstract_declarator) <case DECLTYPE_TYPE>:
New no-op case.

gcc/testsuite/ChangeLog:

PR c++/85278
* g++.dg/concepts/diagnostic9.C: New test.

4 years agoipa-sra: Fix treatment of internal functions (PR 94434)
Martin Jambor [Tue, 14 Apr 2020 17:26:04 +0000 (19:26 +0200)]
ipa-sra: Fix treatment of internal functions (PR 94434)

IPA-SRA can segfault when processing a call to an internal function
because such calls do not have corresponding call graphs and should be
treated like memory accesses anyway, which what the patch below does.

The patch makes an attempt to differentiate between reads and writes,
although for things passed by value it does not make any difference.
It treats all arguments of functions denoted with internal_store_fn_p
as written to, even though in practice only some are, but for IPA-SRA
purposes, actions needed to be taken when processing a read are also
always performed when analyzing a write, so the code is just slightly
pessimistic.  But not as pessimistic as bailing out on any internal
call immediately.

I have LTO bootstrapped and tested the patch on x86_64-linux and
normally bootstrapped and tested it on aarch64-linux, although one
which is not SVE capable.  I would appreciate testing on such machine
too - as well as a small testcase that would follow all relevant
conventions in gcc.target/aarch64/sve.

2020-04-14  Martin Jambor  <mjambor@suse.cz>

PR ipa/94434
* ipa-sra.c: Include internal-fn.h.
(enum isra_scan_context): Update comment.
(scan_function): Treat calls to internal_functions like loads or stores.

4 years agoc++: Stray RESULT_DECLs in result of constexpr call [PR94034]
Patrick Palka [Wed, 8 Apr 2020 17:14:42 +0000 (13:14 -0400)]
c++: Stray RESULT_DECLs in result of constexpr call [PR94034]

When evaluating the initializer of 'a' in the following example

  struct A {
    A() = default; A(const A&);
    A *p = this;
  };
  constexpr A foo() { return {}; }
  constexpr A a = foo();

the PLACEHOLDER_EXPR for 'this' in the aggregate initializer returned by foo
gets resolved to the RESULT_DECL of foo.  But due to guaranteed RVO, the 'this'
should really be resolved to '&a'.

Fixing this properly by immediately resolving 'this' and PLACEHOLDER_EXPRs to
the ultimate object under construction would in general mean that we would no
longer be able to cache constexpr calls for which RVO possibly applies, because
the result of the call may now depend on the ultimate object under construction.

So as a mostly correct stopgap solution that retains cachability of RVO'd
constexpr calls, this patch fixes this issue by rewriting all occurrences of the
RESULT_DECL in the result of a constexpr function call with the current object
under construction, after the call returns.  This means the 'this' pointer
during construction of the temporary will still point to the temporary object
instead of the ultimate object, but besides that this approach seems
functionally equivalent to the proper approach.

gcc/cp/ChangeLog:

PR c++/94034
* constexpr.c (replace_result_decl_data): New struct.
(replace_result_decl_data_r): New function.
(replace_result_decl): New function.
(cxx_eval_call_expression): Use it.
* tree.c (build_aggr_init_expr): Set the location of the AGGR_INIT_EXPR
to that of its initializer.

gcc/testsuite/ChangeLog:

PR c++/94034
* g++.dg/cpp0x/constexpr-empty15.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi6a.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi6b.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi7a.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi7b.C: New test.

4 years agotestsuite: Add testcase for already fixed PR [PR94573]
Jakub Jelinek [Tue, 14 Apr 2020 14:14:47 +0000 (16:14 +0200)]
testsuite: Add testcase for already fixed PR [PR94573]

2020-04-14  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/94573
* gcc.dg/store_merging_30.c: New test.

4 years agoPR tree-optimization/94574 - aarch64: ICE during GIMPLE pass:ccp
Yang Yang [Tue, 14 Apr 2020 19:42:23 +0000 (19:42 +0000)]
PR tree-optimization/94574 - aarch64: ICE during GIMPLE pass:ccp

In this PR the testcase ICEs because a BIT_INSERT_EXPR whose replaced bits are
not fully inside the container is generated. A size check is added to avoid
this kind of ICE.

gcc/ChangeLog:

PR tree-optimization/94574
* tree-ssa.c (non_rewritable_lvalue_p): Add size check when analyzing
whether a vector-insert is rewritable using a BIT_INSERT_EXPR.

gcc/testsuite/ChangeLog:

PR tree-optimization/94574
* gcc.dg/pr94574.c: New test.

4 years agoi386: Remove mode size check in ix86_get_ssemov
H.J. Lu [Tue, 14 Apr 2020 13:00:21 +0000 (06:00 -0700)]
i386: Remove mode size check in ix86_get_ssemov

Even though ix86_hard_regno_mode_ok doesn't allow xmm16-xmm31 nor
ymm16-ymm31 in 128/256 bit modes when AVX512VL is disabled, LRA can
still generate reg to reg moves with xmm16-xmm31 and ymm16-ymm31 in
128/256 bit modes.  Remove mode size check in ix86_get_ssemov.

gcc/

PR target/94561
* config/i386/i386.c (ix86_get_ssemov): Remove mode size check.

gcc/testsuite/

PR target/94561
* gcc.target/i386/pr94561.c: New test.

4 years agoAdd myself as callgraph (IPA) reviewer
Martin Jambor [Tue, 14 Apr 2020 12:12:18 +0000 (14:12 +0200)]
Add myself as callgraph (IPA) reviewer

2020-04-14  Martin Jambor  <mjambor@suse.cz>

* MAINTAINERS (Reviewers): Add myself as callgraph (IPA) reviewer.

4 years ago[Fortran] Fix name conflict check for use-assoc (PR 92736)
Tobias Burnus [Tue, 14 Apr 2020 11:57:52 +0000 (13:57 +0200)]
[Fortran] Fix name conflict check for use-assoc (PR 92736)

* module.c (gfc_match_use): Fix name-conflict check for use-associating
the same symbol again in a submodule.

* gfortran.dg/use_rename_10.f90: New.
* gfortran.dg/use_rename_11.f90: New.

4 years agoFix PR 94270 by not warning about artifical dummy arguments.
Thomas König [Tue, 14 Apr 2020 11:50:51 +0000 (13:50 +0200)]
Fix PR 94270 by not warning about artifical dummy arguments.

2020-04-14  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/94270
* interface.c (gfc_get_formal_from_actual_arglist): Always
set artificial attribute for symbols.
* trans-decl.c (generate_local_decl): Do not warn if the
symbol is artifical.

2020-04-14  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/94270
* gfortran.dg/warn_unused_dummy_argument_6.f90: New test.

4 years agoAdd me as callgraph (IPA) reviewer.
Martin Liska [Tue, 14 Apr 2020 11:19:21 +0000 (13:19 +0200)]
Add me as callgraph (IPA) reviewer.

ChangeLog:

2020-04-14  Martin Liska  <mliska@suse.cz>

* MAINTAINERS: Add me as callgraph (IPA) reviewer.

4 years agolibstdc++: Make comparison category comparisons noexcept (PR 94565)
Jonathan Wakely [Tue, 14 Apr 2020 10:23:34 +0000 (11:23 +0100)]
libstdc++: Make comparison category comparisons noexcept (PR 94565)

PR libstdc++/94565
* libsupc++/compare (__unspec): Add noexcept-specifier to constructor.
* testsuite/18_support/comparisons/categories/94565.cc: New test.

4 years agoc++: Improve redeclared parameter name diagnostic [PR94588]
Marek Polacek [Mon, 13 Apr 2020 23:06:39 +0000 (19:06 -0400)]
c++: Improve redeclared parameter name diagnostic [PR94588]

While reviewing [basic.scope.param] I noticed we don't show the location
of the previous declaration when giving an error about "A parameter name
shall not be redeclared in the outermost block of the function definition".

PR c++/94588
* name-lookup.c (check_local_shadow): Add an inform call.

* g++.dg/diagnostic/redeclaration-1.C: Add dg-message.

4 years agoDaily bump.
GCC Administrator [Tue, 14 Apr 2020 00:16:25 +0000 (00:16 +0000)]
Daily bump.

4 years agoUpdate and correct documentation of -Wall and -Wrestrict.
Martin Sebor [Mon, 13 Apr 2020 22:31:09 +0000 (16:31 -0600)]
Update and correct documentation of -Wall and -Wrestrict.

gcc/ChangeLog:
* doc/extend.texi (-Wall): Mention -Wformat-overflow and
-Wformat-truncation.  Move -Wzero-length-bounds last.
(-Wrestrict): Document positive form of option enabled by -Wall.

4 years agoPR c/92326 - wrong bound in zero-length array diagnostics
Martin Sebor [Mon, 13 Apr 2020 21:31:44 +0000 (15:31 -0600)]
PR c/92326 - wrong bound in zero-length array diagnostics

gcc/c-family/ChangeLog:

PR c/92326
* c-pretty-print.c (c_pretty_printer::direct_abstract_declarator): Avoid
printing array bound for flexible array members.

gcc/testsuite/ChangeLog:

PR c/92326
* c-c++-common/Warray-bounds-8.c: New test.
* gcc.dg/Warray-bounds-46.c: Adjust expected format of flexible array
memebrs in diagnostics.
* gcc.dg/Warray-bounds-49.c: Same.

4 years agoc++: Infinite diagnostic loop with decltype([]{}) [PR94521]
Patrick Palka [Mon, 13 Apr 2020 20:53:41 +0000 (16:53 -0400)]
c++: Infinite diagnostic loop with decltype([]{}) [PR94521]

We are hitting a recursive loop when printing the signature of a function
containing a decltype([]{}) type.  The loop is

  dump_function_decl -> dump_substitution
     -> dump_template_bindings
     -> dump_type
     -> dump_aggr_type
     -> dump_scope -> dump_function_decl

and we loop because dump_template_bindings wants to print the resolved type of
decltype([]{}) (i.e. just a lambda type), so it calls dump_aggr_type, which
wants to print the function scope of the lambda type.  But the function scope of
the lambda type is the function which we're in the middle of printing.

This patch breaks the loop by passing TFF_NO_FUNCTION_ARGUMENTS to
dump_function_decl from dump_scope, so that we avoid recursing into
dump_substitution and ultimately looping.

This also means we no longer emit the "[with ...]" clause when printing a
function template scope, and we instead just emit its template argument list in
a more natural way, e.g. instead of
    foo(int, char) [with T=bool]::x
we would now print
    foo<bool>::x
which seems like an improvement on its own.

The full signature of the function 'spam' in the below testcase is now
  void spam(decltype (<lambda>)*) [with T = int; decltype (<lambda>) = spam<int>::<lambda()>]

gcc/cp/ChangeLog:

PR c++/94521
* error.c (dump_scope): Pass TFF_NO_FUNCTION_ARGUMENTS to
dump_function_decl when printing a function template instantiation as a
scope.

gcc/testsuite/ChangeLog:

PR c++/94521
* g++.dg/cpp2a/lambda-uneval12.C: New test.

4 years agoc++: More self-modifying constexpr init [PR94470]
Patrick Palka [Mon, 13 Apr 2020 20:53:02 +0000 (16:53 -0400)]
c++: More self-modifying constexpr init [PR94470]

In this PR we're incorrectly rejecting a self-modifying constexpr initializer as
a consequence of the fix for PR78572.

It looks like however that the fix for PR78572 is obsoleted by the fix for
PR89336: the testcase from the former PR successfully compiles even with its fix
reverted.

But then further testing showed that the analogous testcase of PR78572 where the
array has an aggregate element type is still problematic (i.e. we ICE) even with
the fix for PR78572 applied.  The reason is that in cxx_eval_bare_aggregate we
attach a constructor_elt of aggregate type always to the end of the new
CONSTRUCTOR, but that's not necessarily correct if the CONSTRUCTOR is
self-modifying.  We should instead be using get_or_insert_ctor_field to insert
the constructor_elt in the right place.

So this patch reverts the PR78572 fix and makes the appropriate changes to
cxx_eval_bare_aggregate.  This fixes PR94470, and we now are also able to fully
reduce the initializers of 'arr' and 'arr2' in the new test array57.C to
constant initializers.

gcc/cp/ChangeLog:

PR c++/94470
* constexpr.c (get_or_insert_ctor_field): Set default value of parameter
'pos_hint' to -1.
(cxx_eval_bare_aggregate): Use get_or_insert_ctor_field instead of
assuming the the next index belongs at the end of the new CONSTRUCTOR.
(cxx_eval_store_expression): Revert PR c++/78572 fix.

gcc/testsuite/ChangeLog:

PR c++/94470
* g++.dg/cpp1y/constexpr-nsdmi8.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi9.C: New test.
* g++.dg/init/array57.C: New test.

4 years agocoroutines: Rename the coroutines cpp builtin.
Iain Sandoe [Mon, 13 Apr 2020 19:48:46 +0000 (20:48 +0100)]
coroutines: Rename the coroutines cpp builtin.

The current standard draft (n4861) amends the cpp builtin for
coroutines to '__cpp_impl_coroutine', no other change.

gcc/c-family/ChangeLog:

2020-04-13  Iain Sandoe  <iain@sandoe.co.uk>

* c-cppbuiltin.c (c_cpp_builtins): Update coroutines builtin
define, per n4861.

gcc/testsuite/ChangeLog:

2020-04-13  Iain Sandoe  <iain@sandoe.co.uk>

* g++.dg/coroutines/coro-pre-proc.C: Update coroutines builtin
define, per n4861.
* g++.dg/coroutines/coro.h: Likewise.

libstdc++-v3/ChangeLog:

2020-04-13  Iain Sandoe  <iain@sandoe.co.uk>

* include/std/coroutine: Update coroutines builtin define,
per n4861.

4 years agoMinor documentation fix
Zackery Spytz [Mon, 13 Apr 2020 19:29:13 +0000 (13:29 -0600)]
Minor documentation fix

* doc/extend.texi: Add realloc to list of built-in functions
are recognized by the compiler.

4 years agoDarwin, testsuite: Fix darwin-version-1.c fails with XCode 11.4.
Iain Sandoe [Mon, 13 Apr 2020 18:27:21 +0000 (19:27 +0100)]
Darwin, testsuite: Fix darwin-version-1.c fails with XCode 11.4.

From XCode 11.4 on 10.14/15 use of 10.6 and 10.7 is deprecated.
The tools issue diagnostics if -mmacosx-version-min= < 10.8

Adjust the testcase to avoid that usage on 10.14, 10.15 for now.

gcc/testsuite/ChangeLog:

2020-04-13  Iain Sandoe  <iain@sandoe.co.uk>

* gcc.dg/darwin-version-1.c: Use -mmacosx-version-min= 10.8
for system versions 10.14 and 10.15.

4 years agoCommit test case for PR 87644, which has fixed itself somehow.
Thomas König [Mon, 13 Apr 2020 17:06:41 +0000 (19:06 +0200)]
Commit test case for PR 87644, which has fixed itself somehow.

2020-04-13  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/87644
* gfortran.dg/variable_parameter.f90: New test.

4 years agoICE on wrong code [PR94192].
Linus Koenig [Mon, 13 Apr 2020 14:30:44 +0000 (16:30 +0200)]
ICE on wrong code [PR94192].

The idea is not have another resolution of a pointer if an error has
occurred previously.

2020-04-13  Linus Koenig <link@sig-st.de>

PR fortran/94192
* resolve.c (resolve_fl_var_and_proc): Set flag "error" to 1 if
pointer is found to not have an assumed rank or a deferred shape.
* simplify.c (simplify_bound): If an error has been issued for a
given pointer, one should not attempt to find its bounds.

2020-04-13  Linus Koenig <link@sig-st.de>

PR fortran/94192
* gfortran.dg/bound_resolve_after_error_1.f90: New test.

4 years agoc++: lambdas with internal linkage are different to no-linkage [PR94426]
Nathan Sidwell [Mon, 13 Apr 2020 13:35:33 +0000 (06:35 -0700)]
c++: lambdas with internal linkage are different to no-linkage [PR94426]

My fix for 94147 was confusing no-linkage with internal linkage, at
the language level.  That's wrong. (the std is confusing here, because
it describes linkage of names (which is wrong), and lambdas have no
names)

Lambdas with extra-scope, have linkage.  However, at the
implementation-level that linkage is at least as restricted as the
linkage of the extra-scope decl.

Further, when instantiating a variable initialized by a lambda, we
must determine the visibility of the variable itself, before
instantiating its initializer.  If the template arguments are internal
(or no-linkage), the variable will have internal linkage, regardless
of the linkage of the template it is instantiated from.  We need to
know that before instantiating the lambda, so we can restrict its
linkage correctly.

* decl2.c (determine_visibility): A lambda's visibility is
affected by its extra scope.
* pt.c (instantiate_decl): Determine var's visibility before
instantiating its initializer.
* tree.c (no_linkage_check): Revert code looking at visibility of
lambda's extra scope.
` gcc/cp/
* g++.dg/cpp0x/lambda/pr94426-[12].C: New.
* g++.dg/abi/lambda-vis.C: Drop a warning.
* g++.dg/cpp0x/lambda/lambda-mangle.C: Lambda visibility on
variable changes.
* g++.dg/opt/dump1.C: Drop warnings of no import.

4 years agox86: Restore the frame pointer in word_mode
H.J. Lu [Mon, 13 Apr 2020 11:23:05 +0000 (04:23 -0700)]
x86: Restore the frame pointer in word_mode

We must restore the frame pointer in word_mode for eh_return epilogues
since the upper 32 bits of RBP register can have any values.

Tested on Linux/x32 and Linux/x86-64.

PR target/94556
* config/i386/i386.c (ix86_expand_epilogue): Restore the frame
pointer in word_mode for eh_return epilogues.

4 years agoMSP430: Dont add offsets to addresses when emitting asm for post_inc
Jozef Lawrynowicz [Fri, 10 Apr 2020 16:31:33 +0000 (17:31 +0100)]
MSP430: Dont add offsets to addresses when emitting asm for post_inc

Some insns, which operate on SImode operands, output assembler template
that comprise of multiple instructions using HImode operands. To access
the high word of an SImode operand, an operand selector '%H' is used to
offset the operand value by a constant amount.

When one of these HImode operands is a memory reference to a post_inc,
the address does not need to be offset, since the preceding instruction
has already offset the address to the correct value.

This fixes an ICE in change_address_1, at emit-rtl.c:2318 for
gcc.c-torture/execute/pr20527-1.c in the "-mlarge -O2 -flto
-fuse-linker-plugin -fno-fat-lto-objects" configuration.

This test generated the following insn, and the attempt to output the
high part of the post_inc address caused the ICE.
(set (reg:SI 6 R6)
     (minus:SI (reg:SI 6 R6)
               (mem:SI (post_inc:PSI (reg:PSI 10 R10)) {subsi3}

gcc/ChangeLog:

2020-04-13  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config/msp430/msp430.c (msp430_print_operand): Don't add offsets to
memory references in %B, %C and %D operand selectors when the inner
operand is a post increment address.

4 years agoMSP430: Fix memory offsets used by %C and %D asm output operand modifiers
Jozef Lawrynowicz [Mon, 13 Apr 2020 09:28:01 +0000 (10:28 +0100)]
MSP430: Fix memory offsets used by %C and %D asm output operand modifiers

The %C and %D operand modifiers are supposed to access the 3rd and 4th
words of a 64-bit value, so for memory references they need to offset
the given address by 4 and 6 bytes respectively.

gcc/ChangeLog:

2020-04-13  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* config/msp430/msp430.c (msp430_print_operand): Offset a %C memory
reference by 4 bytes, and %D memory reference by 6 bytes.

gcc/testsuite/ChangeLog:

2020-04-13  Jozef Lawrynowicz  <jozef.l@mittosystems.com>

* gcc.target/msp430/operand-modifiers.c: New test.

4 years agod: Merge update dmd 799066f49
Iain Buclaw [Mon, 13 Apr 2020 09:25:26 +0000 (11:25 +0200)]
d: Merge update dmd 799066f49

Removes the implementation of __traits(argTypes), which only supported
x86_64 targets.  The only use of this trait is an unused va_arg()
function, this has been removed as well.

Reviewed-on: https://github.com/dlang/dmd/pull/11022

gcc/d/ChangeLog:

2020-04-13  Iain Buclaw  <ibuclaw@gdcproject.org>

* Make-lang.in (D_FRONTEND_OBJS): Remove d/argtypes.o.
* d-target.cc (Target::toArgTypes): New function.

libphobos/ChangeLog:

2020-04-13  Iain Buclaw  <ibuclaw@gdcproject.org>

* libdruntime/core/stdc/stdarg.d: Remove run-time va_list template.

4 years agoRename 'libgomp.oacc-c-c++-common/static-dynamic-lifetimes-*' to 'libgomp.oacc-c...
Thomas Schwinge [Mon, 13 Apr 2020 06:56:03 +0000 (08:56 +0200)]
Rename 'libgomp.oacc-c-c++-common/static-dynamic-lifetimes-*' to 'libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-*' [PR92843]

Fix-up for commit be9862dd96945772ae0692bc95b37ec6dbcabda0 "Test cases for
mixed structured/dynamic data lifetimes with OpenACC [PR92843]": it's
"structured", not "static" data lifetimes/reference counters.

libgomp/
PR libgomp/92843
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-1-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-1-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-1.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-1.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-2-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-2-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-2.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-2.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-3-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-3-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-3.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-3.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-4-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-4-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-4.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-4.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-5-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-5-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-5.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-5.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-6-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-6-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-6.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-6.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-7-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-7-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-7.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-7.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-8-lib.c:
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-8-lib.c:
... this.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-8.c::
Rename to...
* testsuite/libgomp.oacc-c-c++-common/structured-dynamic-lifetimes-8.c:
... this.

4 years agoDaily bump.
GCC Administrator [Mon, 13 Apr 2020 00:16:18 +0000 (00:16 +0000)]
Daily bump.

4 years agoAdded a test case from a PR that was fixed by something else.
Thomas König [Sun, 12 Apr 2020 15:02:08 +0000 (17:02 +0200)]
Added a test case from a PR that was fixed by something else.

2020-04-12  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/94091
* gfortran.dg/char_length_22.f90: New test.

Committed ChangeLog entry.

4 years agoDarwin, testsuite: Adjust indirect thunk tests for default flags.
Iain Sandoe [Sun, 12 Apr 2020 12:15:38 +0000 (13:15 +0100)]
Darwin, testsuite: Adjust indirect thunk tests for default flags.

Darwin mandates an indirection for variables in the commmon
section.  Since the change to -fno-common, variables in some
of the thunk tests are now in the .data section where they
may be accessed directly.  Remove the indirections from the
scan-assembler matches.

gcc/testsuite/ChangeLog:

2020-04-12  Iain Sandoe  <iain@sandoe.co.uk>

* gcc.target/i386/indirect-thunk-1.c: Adjust for fno-common
change, removing indirections for vars in .data.
* gcc.target/i386/indirect-thunk-2.c: Likewise.
* gcc.target/i386/indirect-thunk-3.c: Likewise.
* gcc.target/i386/indirect-thunk-4.c: Likewise.
* gcc.target/i386/indirect-thunk-attr-1.c: Likewise.
* gcc.target/i386/indirect-thunk-attr-2.c: Likewise.
* gcc.target/i386/indirect-thunk-attr-3.c: Likewise.
* gcc.target/i386/indirect-thunk-attr-4.c: Likewise.
* gcc.target/i386/indirect-thunk-attr-5.c: Likewise.
* gcc.target/i386/indirect-thunk-attr-6.c: Likewise.
* gcc.target/i386/indirect-thunk-extern-1.c: Likewise.
* gcc.target/i386/indirect-thunk-extern-2.c: Likewise.
* gcc.target/i386/indirect-thunk-extern-3.c: Likewise.
* gcc.target/i386/indirect-thunk-extern-4.c: Likewise.
* gcc.target/i386/indirect-thunk-inline-1.c: Likewise.
* gcc.target/i386/indirect-thunk-inline-2.c: Likewise.
* gcc.target/i386/indirect-thunk-inline-3.c: Likewise.
* gcc.target/i386/indirect-thunk-inline-4.c: Likewise.

4 years agoDaily bump.
GCC Administrator [Sun, 12 Apr 2020 00:16:16 +0000 (00:16 +0000)]
Daily bump.

4 years agoc++: Add test for PR 86327.
Marek Polacek [Sat, 11 Apr 2020 14:45:33 +0000 (10:45 -0400)]
c++: Add test for PR 86327.

Fixed by r264171.

PR c++/86327
* g++.dg/cpp1y/constexpr-return5.C: New test.

4 years agoi386: Fix REDUC_SSE_SMINMAX_MODE mode conditions.
Uros Bizjak [Sat, 11 Apr 2020 09:51:41 +0000 (11:51 +0200)]
i386: Fix REDUC_SSE_SMINMAX_MODE mode conditions.

V4SI, V8HI and V16QI modes of redux_<code>_scal_<mode> expander
expand with SSE2 instructions (PSRLDQ and PCMPGTx) so use
TARGET_SSE2 as relevant mode iterator codition.

PR target/94494
* config/i386/sse.md (REDUC_SSE_SMINMAX_MODE): Use TARGET_SSE2
condition for V4SI, V8HI and V16QI modes.

testsuite/ChangeLog:

PR target/94494
* gcc.target/i386/pr94494.c: New test.

4 years agotestsuite: Fix up pr94482.c testcase [PR94482]
Jakub Jelinek [Sat, 11 Apr 2020 05:50:50 +0000 (07:50 +0200)]
testsuite: Fix up pr94482.c testcase [PR94482]

The test FAILs on powerpc64-linux with -m32 due to psabi warnings.
Furthermore, the test needs really -msse2 to reproduce on x86 -m32 at -O2.

2020-04-11  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/94482
* gcc.dg/torture/pr94482.c: Add -Wno-psabi -w.  Don't add -msse
and sse_runtime effective target on x86, instead only add -msse2
if target is sse2_runtime.

4 years agocselib: Mark the cselib_record_sp_cfa_base_equiv VALUE as preserved [PR94551]
Jakub Jelinek [Sat, 11 Apr 2020 05:32:12 +0000 (07:32 +0200)]
cselib: Mark the cselib_record_sp_cfa_base_equiv VALUE as preserved [PR94551]

Sometimes the cselib_record_sp_cfa_base_equiv makes it into the var-tracking
used VALUEs and needs to be preserved.

2020-04-11  Jakub Jelinek  <jakub@redhat.com>

PR debug/94495
PR target/94551
* cselib.c (cselib_record_sp_cfa_base_equiv): Set PRESERVED_VALUE_P on
val->val_rtx.

4 years agoDaily bump.
GCC Administrator [Sat, 11 Apr 2020 00:16:15 +0000 (00:16 +0000)]
Daily bump.

4 years agoCorrect PR numbers in the last Changelog.
Iain Sandoe [Fri, 10 Apr 2020 23:46:52 +0000 (00:46 +0100)]
Correct PR numbers in the last Changelog.

I mistyped the PR number in the actual fix commit
r10-7682-g0666767eb4cc864f00ba34d97b9d58f8dc650bba.

4 years agocoroutines: Revise await expansions [PR94528]
Iain Sandoe [Fri, 10 Apr 2020 19:55:10 +0000 (20:55 +0100)]
coroutines: Revise await expansions [PR94528]

The expansions for await expressions were specific to particular
cases, this revises it to be more generic.

a: Revise co_await statement walkers.

We want to process the co_awaits one statement at a time.
We also want to be able to determine the insertion points for
new bind scopes needed to cater for temporaries that are
captured by reference and have lifetimes that need extension
to the end of the full expression.  Likewise, the handling of
captured references in the evaluation of conditions might
result in the need to make a frame copy.

This reorganises the statement walking code to make it easier to
extend for these purposes.

b: Factor reference-captured temp code.

We want to be able to use the code that writes a new bind expr
with vars (and their initializers) from several places, so split
that out of the maybe_promote_captured_temps() function into a
new replace_statement_captures ().  Update some comments.

c: Generalize await statement expansion.

This revises the expansion to avoid the need to expand conditionally
on the tree type.  It resolves PR 94528.

gcc/cp/ChangeLog:

2020-04-10  Iain Sandoe  <iain@sandoe.co.uk>

PR c++/94538
* coroutines.cc (co_await_expander): Remove.
(expand_one_await_expression): New.
(process_one_statement): New.
(await_statement_expander): New.
(build_actor_fn): Revise to use per-statement expander.
(struct susp_frame_data): Reorder and comment.
(register_awaits): Factor code.
(replace_statement_captures): New, factored from...
(maybe_promote_captured_temps):.. here.
(await_statement_walker): Revise to process per statement.
(morph_fn_to_coro): Use revised susp_frame_data layout.

gcc/testsuite/ChangeLog:

2020-04-10  Iain Sandoe  <iain@sandoe.co.uk>

PR c++/94538
* g++.dg/coroutines/pr94528.C: New test.

4 years agoc++: make __is_constructible work with paren-init of aggrs [PR94149]
Marek Polacek [Thu, 9 Apr 2020 20:31:59 +0000 (16:31 -0400)]
c++: make __is_constructible work with paren-init of aggrs [PR94149]

In C++20 this is well-formed:

  using T = int[2];
  T t(1, 2);

which means that std::is_constructible_v<int[2], int, int> should be true.
But constructible_expr immediately returned the error_mark_node when it
saw a list with more than one element.  To give accurate results in
C++20, we have to try initializing the aggregate from a parenthesized list of
values.

To not repeat the same mistake as in c++/93790, if there's only one
element, I'm trying {} only when () didn't succeed.  is_constructible5.C
verifies this.

In paren-init24.C std::is_nothrow_constructible_v doesn't work due to
 error: invalid 'static_cast' from type 'int' to type 'int [1]'
and
 error: functional cast to array type 'int [2]'

This needs to be fixed in libstdc++.

PR c++/94149
* method.c (constructible_expr): In C++20, try using parenthesized
initialization of aggregates to determine the result of
__is_constructible.

* g++.dg/cpp2a/paren-init24.C: New test.
* g++.dg/cpp2a/paren-init25.C: New test.
* g++.dg/ext/is_constructible5.C: New test.

4 years agoFix typo in gfortran.dg/asynchronous_5.f03 from last commit.
Fritz Reese [Fri, 10 Apr 2020 15:28:45 +0000 (11:28 -0400)]
Fix typo in gfortran.dg/asynchronous_5.f03 from last commit.

2020-04-10  Fritz Reese  <foreese@gcc.gnu.org>

* gfortran.dg/asynchronous_5.f03: Fix typo in testcase and add
IMPLICIT NONE.

4 years agoFix UNRESOLVED testcase gfortran.dg/asynchronous_5.f03.
Fritz Reese [Fri, 10 Apr 2020 15:24:28 +0000 (11:24 -0400)]
Fix UNRESOLVED testcase gfortran.dg/asynchronous_5.f03.

gcc/testsuite/ChangeLog:

2020-04-10  Fritz Reese  <foreese@gcc.gnu.org>

* gfortran.dg/asynchronous_5.f03: Add -fdump-tree-original and fix
patterns for scan-tree-dump.

4 years agoTest cases for mixed structured/dynamic data lifetimes with OpenACC [PR92843]
Julian Brown [Fri, 17 Jan 2020 21:18:18 +0000 (13:18 -0800)]
Test cases for mixed structured/dynamic data lifetimes with OpenACC [PR92843]

libgomp/
PR libgomp/92843
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-1-lib.c:
New file.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-1.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-2-lib.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-2.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-3-lib.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-3.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-4-lib.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-4.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-5-lib.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-5.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-6-lib.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-6.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-7-lib.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-7.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-8-lib.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/static-dynamic-lifetimes-8.c:
Likewise.

4 years agoAdd 'dg-do run' to 'libgomp.fortran/target-enter-data-1.f90'
Thomas Schwinge [Wed, 1 Apr 2020 21:26:56 +0000 (23:26 +0200)]
Add 'dg-do run' to 'libgomp.fortran/target-enter-data-1.f90'

Fix-up for commit 689418b97e5eb6a221871a2439bca3e6283ac579 "libgomp – fix
handling of 'target enter data'".

libgomp/
* testsuite/libgomp.fortran/target-enter-data-1.f90: Add 'dg-do
run'.