platform/upstream/gcc.git
3 years agoMAINTAINERS updates for ex-ImgTec employees
Jeff Law [Thu, 11 Mar 2021 16:28:50 +0000 (09:28 -0700)]
MAINTAINERS updates for ex-ImgTec employees

/
* MAINTAINERS: Update entries for a few ex-ImgTec employees

3 years agoAdd -fprofile-reproducible=parallel-runs to STAGEfeedback_CFLAGS to Makefile.tpl.
Martin Liska [Thu, 11 Mar 2021 15:17:41 +0000 (16:17 +0100)]
Add -fprofile-reproducible=parallel-runs to STAGEfeedback_CFLAGS to Makefile.tpl.

ChangeLog:

* Makefile.tpl: The change was done Makefile.in which
is generated file.

3 years agotestsuite/98245 - adjust dump scanning of gcc.dg/vect/bb-slp-46.c
Richard Biener [Thu, 11 Mar 2021 13:25:16 +0000 (14:25 +0100)]
testsuite/98245 - adjust dump scanning of gcc.dg/vect/bb-slp-46.c

Checking the number of pluses is unreliable since the vector size
isn't known.  Instead see that the unwanted scalar compute is not
there.

2021-03-11  Richard Biener  <rguenther@suse.de>

PR testsuite/98245
* gcc.dg/vect/bb-slp-46.c: Scan for the scalar compute
instead of verifying the total number of adds.

3 years agotestsuite/97494 - XFAIL gcc.dg/vect/pr97428.c on !vect_hw_misalign
Richard Biener [Thu, 11 Mar 2021 13:11:08 +0000 (14:11 +0100)]
testsuite/97494 - XFAIL gcc.dg/vect/pr97428.c on !vect_hw_misalign

While we could at least vectorize it on targets which support
re-alignment tokens we fail to do this because of imperfections in
alignment analysis.  XFAIL when the HW cannot deal with misaligned
vector accesses for now.

2021-03-11  Richard Biener  <rguenther@suse.de>

PR testsuite/97494
* gcc.dg/vect/pr97428.c: XFAIL on !vect_hw_misalign.

3 years agotestsuite/97494 - XFAIL gcc.dg/vect/vect-complex-5.c on !vect_hw_misalign
Richard Biener [Thu, 11 Mar 2021 13:03:55 +0000 (14:03 +0100)]
testsuite/97494 - XFAIL gcc.dg/vect/vect-complex-5.c on !vect_hw_misalign

This is a missed optimization due to bogus alignment analysis.

2021-03-11  Richard Biener  <rguenther@suse.de>

PR testsuite/97494
* gcc.dg/vect/vect-complex-5.c: XFAIL on !vect_hw_misalign.

3 years agotestsuite/97494 - amend gcc.dg/vect/slp-21.c
Richard Biener [Thu, 11 Mar 2021 12:30:56 +0000 (13:30 +0100)]
testsuite/97494 - amend gcc.dg/vect/slp-21.c

As reported in the PR all powerpc64 targets fail

FAIL: gcc.dg/vect/slp-21.c scan-tree-dump-times vect "vectorizing stmts using SLP" 2

because like on arm we now vectorize 4 opportunities.  This adjusts
the testcase to follow the arm example.

2021-03-11  Richard Biener  <rguenther@suse.de>

PR testsuite/97494
* gcc.dg/vect/slp-21.c: Adjust for powerpc64*-*-*.

3 years agotree-optimization/99523 - missing SSA decls in dumps
Richard Biener [Thu, 11 Mar 2021 09:09:43 +0000 (10:09 +0100)]
tree-optimization/99523 - missing SSA decls in dumps

This makes sure to dump SSA names without identifier in the
declaration part of a function dump.  While we dump the
anonymous variable decls the SSA names referencing them appear
without a clear reference as to what anonymous variable is used
(_3 vs. D.1234).

2021-03-11  Richard Biener  <rguenther@suse.de>

PR tree-optimization/99523
* tree-cfg.c (dump_function_to_file): Dump SSA names
w/o identifier to the decls section as well, not only those
without a VAR_DECL.

3 years agoicf: Check return type of internal fn calls [PR99517]
Jakub Jelinek [Thu, 11 Mar 2021 09:59:18 +0000 (10:59 +0100)]
icf: Check return type of internal fn calls [PR99517]

The following testcase is miscompiled, because IPA-ICF considers the two
functions identical.  They aren't, the types of the .VEC_CONVERT call
lhs is different.  But for calls to internal functions, there is no
fntype nor callee with a function type to compare, so all we compare
is just the ifn, arguments and some call flags.

The following patch fixes it by checking the internal fn calls like e.g. gimple
assignments where the type of the lhs is checked too.

2021-03-11  Jakub Jelinek  <jakub@redhat.com>

PR ipa/99517
* ipa-icf-gimple.c (func_checker::compare_gimple_call): For internal
function calls with lhs fail if the lhs don't have compatible types.

* gcc.target/i386/avx2-pr99517-1.c: New test.
* gcc.target/i386/avx2-pr99517-2.c: New test.

3 years agocris: define HARD_FRAME_POINTER_REGNUM
Hans-Peter Nilsson [Thu, 11 Mar 2021 02:38:34 +0000 (03:38 +0100)]
cris: define HARD_FRAME_POINTER_REGNUM

Beware, tm.texi doesn't tell the whole story: a defined
HARD_FRAME_POINTER_REGNUM (different to FRAME_POINTER_REGNUM) is
supposed to make work easier for reload, being able to easily
tell actual frame-pointer-related addresses from those that
happen to use the same register or something to that effect.

On reasonable code the performance effect is barely measurable.

Looking at libgcc changes for -march=v10, the effect (where
noticeable) is mostly indeterminate churn.  Instances where it's
not just insns moved around at no obvious effect: one more insn
for addvdi3, subvdi3; two insns more in floatdisf; three insns
shorter fixunsdfdi.  Some of those seem related to pairing r8
with r9.  The only effect on coremark is an infinitesimal
positive effect from a three(!) cycles total (from the 15 calls)
faster execution paths in vfprintf_r.  Local microbenchmarks
give similar results.  With that in mind and not forgetting that
expectations in the register allocator and reload leaning
towards HARD_FRAME_POINTER_REGNUM defined (and different to)
FRAME_POINTER_REGNUM or to wit, "all the kids do it", why not.
Note that the offset at elimination really is 0.

gcc:
* config/cris/cris.h (HARD_FRAME_POINTER_REGNUM): Define.
Change FRAME_POINTER_REGNUM to correspond to a new faked
register faked_fp, part of GENNONACR_REGS like faked_ap.
(CRIS_FAKED_REGS_CONTENTS): New helper macro.
(FIRST_PSEUDO_REGISTER, FIXED_REGISTERS, CALL_USED_REGISTERS):
(REG_ALLOC_ORDER, REG_CLASS_CONTENTS, REGNO_OK_FOR_BASE_P)
(ELIMINABLE_REGS, REGISTER_NAMES): Adjust accordingly.
* config/cris/cris.md (CRIS_FP_REGNUM): Renumber to new faked
register.
(CRIS_REAL_FP_REGNUM): New constant.
* config/cris/cris.c (cris_reg_saved_in_regsave_area): Check
for HARD_FRAME_POINTER_REGNUM instead of FRAME_POINTER_REGNUM.
(cris_initial_elimination_offset): Handle elimination changes
to HARD_FRAME_POINTER_REGNUM instead of FRAME_POINTER_REGNUM
and add one from FRAME_POINTER_REGNUM to
HARD_FRAME_POINTER_REGNUM.
(cris_expand_prologue, cris_expand_epilogue): Emit code for
hard_frame_pointer_rtx instead of frame_pointer_rtx.

3 years agoDaily bump.
GCC Administrator [Thu, 11 Mar 2021 00:16:37 +0000 (00:16 +0000)]
Daily bump.

3 years agoaix: align double complex
David Edelsohn [Tue, 9 Mar 2021 22:52:36 +0000 (17:52 -0500)]
aix: align double complex

AIX word-aligns floating point doubles.  This behavior also extends to
double _Complex, which had been overlooked when compiler support for
double _Complex was added.

This patch adds DCmode to the modes whose alignment is adjusted and
adds a testcase to confirm the correct alignment.

gcc/ChangeLog:

2021-03-10  David Edelsohn  <dje.gcc@gmail.com>

PR target/99492
* config/rs6000/aix.h (ADJUST_FIELD_ALIGN): Add check for DCmode.
* config/rs6000/rs6000.c (rs6000_special_round_type_align): Same.

gcc/testsuite/ChangeLog:

2021-03-10  David Edelsohn  <dje.gcc@gmail.com>

PR target/99492
* gcc.target/powerpc/pr99492.c: New testcase.

3 years agoPR fortran/99205 - Out of memory with undefined character length
Harald Anlauf [Wed, 10 Mar 2021 21:59:50 +0000 (22:59 +0100)]
PR fortran/99205 - Out of memory with undefined character length

A character variable appearing as a data statement object cannot
be automatic, thus it shall have constant length.

gcc/fortran/ChangeLog:

PR fortran/99205
* data.c (gfc_assign_data_value): Reject non-constant character
length for lvalue.
* trans-array.c (gfc_conv_array_initializer): Restrict loop to
elements which are defined to avoid NULL pointer dereference.

gcc/testsuite/ChangeLog:

PR fortran/99205
* gfortran.dg/data_char_4.f90: New test.
* gfortran.dg/data_char_5.f90: New test.

3 years ago[PR99422] LRA: Don't check unknown constraint, use X for empty constraint
Vladimir N. Makarov [Wed, 10 Mar 2021 21:15:08 +0000 (16:15 -0500)]
[PR99422] LRA: Don't check unknown constraint, use X for empty constraint

Using CONSTRAINT__UNKNOWN was a bad idea, although it triggered a lot
hidden bugs.  It is better to use X instead of empty constraint.

gcc/ChangeLog:

PR target/99422
* lra-constraints.c (process_address_1): Don't check unknown
constraint, use X for empty constraint.

3 years agors6000: Fix pr98959 testcase
Segher Boessenkool [Wed, 10 Mar 2021 20:19:39 +0000 (20:19 +0000)]
rs6000: Fix pr98959 testcase

It needs the int128 selector because it uses __int128, and the lp64
selector is the best we can do for -mcmodel=.

2021-03-10  Segher Boessenkool  <segher@kernel.crashing.org>

gcc/testsuite/
* gcc.target/powerpc/pr98959.c: Add int128 and lp64 selectors.

3 years agoFortran: Fix wording in intrinsic.texi's MIN
Tobias Burnus [Wed, 10 Mar 2021 20:24:18 +0000 (21:24 +0100)]
Fortran: Fix wording in intrinsic.texi's MIN

gcc/fortran/ChangeLog:

* intrinsic.texi (MIN): Correct 'maximum' to 'minimum'.

3 years agoc++: ICE do to GC leakage [PR 99423]
Nathan Sidwell [Wed, 10 Mar 2021 20:07:24 +0000 (12:07 -0800)]
c++: ICE do to GC leakage [PR 99423]

My reworking of pending-entity loading introduced a GC problem.  The
post-load processing needs to inhibit GCs (that would otherwise occur
in clone_decl).  That wasn't happening on one code path, leading to
dangling pointers in the active call frames.

PR c++/99423
gcc/cp/
* module.cc (post_load_processing): Assert not gcable.
(laxy_load_pendings): Extend no-gc region around
post_load_processing.
gcc/testsuite/
* g++.dg/modules/pr99423_a.H: New.
* g++.dg/modules/pr99423_b.H: New.

3 years agoanalyzer: factor out new class feasibility_state
David Malcolm [Wed, 10 Mar 2021 17:02:07 +0000 (12:02 -0500)]
analyzer: factor out new class feasibility_state

As preparatory work for a fix to PR analyzer/96374, this patch
moves the core state-update logic from the loop in
exploded_path::feasible_p into a new class feasibility_state.

No functional change intended.

gcc/analyzer/ChangeLog:
PR analyzer/96374
* engine.cc (exploded_path::feasible_p): Move "snodes_visited" and
"model" locals into a new class feasibility_state.  Move heart
of per-edge processing into
feasibility_state::maybe_update_for_edge.
(feasibility_state::feasibility_state): New.
(feasibility_state::maybe_update_for_edge): New, based on loop
body in exploded_path::feasible_p.
* exploded-graph.h (class feasibility_state): New.

3 years agocommitted: analyzer: remove duplicated vfuncs
David Malcolm [Wed, 10 Mar 2021 17:00:34 +0000 (12:00 -0500)]
committed: analyzer: remove duplicated vfuncs

Implement dyn_cast_callgraph_superedge once in callgraph_superedge,
rather than twice in the two subclasses.

Spotted whilst working on a patch for call summaries.

gcc/analyzer/ChangeLog:
* supergraph.h
(callgraph_superedge::dyn_cast_callgraph_superedge): New.
(call_superedge::dyn_cast_callgraph_superedge): Delete.
(return_superedge::dyn_cast_callgraph_superedge): Delete.

3 years agotestsuite: Fix up pr99305.C test on unsigned_char targets [PR99498]
Jakub Jelinek [Wed, 10 Mar 2021 16:40:25 +0000 (17:40 +0100)]
testsuite: Fix up pr99305.C test on unsigned_char targets [PR99498]

On unsigned_char targets, the cast stmt to unsigned char is obviously
not needed (and shouldn't be there).  But it doesn't hurt to test
the rest also on targets where char is unsigned.

2021-03-10  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/99305
PR testsuite/99498
* g++.dg/opt/pr99305.C: Don't expect cast to unsigned char on
unsigned_char effective targets.

3 years agoAdd options to link with libatomic in various tests.
John David Anglin [Wed, 10 Mar 2021 16:01:12 +0000 (16:01 +0000)]
Add options to link with libatomic in various tests.

libstdc++-v3/ChangeLog:

* testsuite/29_atomics/atomic/wait_notify/bool.cc: Add options to
link with libatomic.
* testsuite/29_atomics/atomic/wait_notify/generic.cc: Likewise.
* testsuite/29_atomics/atomic/wait_notify/pointers.cc: Likewise.
* testsuite/29_atomics/atomic_flag/wait_notify/1.cc: Likewise.
* testsuite/30_threads/barrier/arrive.cc: Likewise.
* testsuite/30_threads/barrier/arrive_and_drop.cc: Likewise.
* testsuite/30_threads/barrier/arrive_and_wait.cc: Likewise.
* testsuite/30_threads/barrier/completion.cc: Likewise.
* testsuite/30_threads/latch/3.cc: Likewise.
* testsuite/30_threads/semaphore/try_acquire.cc: Likewise.
* testsuite/30_threads/semaphore/try_acquire_for.cc: Likewise.
* testsuite/30_threads/semaphore/try_acquire_until.cc: Likewise.

3 years agoc++: Propagate assembler name from local-externs [PR 99508]
Nathan Sidwell [Wed, 10 Mar 2021 15:39:51 +0000 (07:39 -0800)]
c++: Propagate assembler name from local-externs [PR 99508]

This is another place where our one-true-decl representation breaks
down.  The fix here propagates the assembly name to the ns-scope
alias.  that fixes the reported problem but changes the behaviour when
the user has explicitly declared the entity in its namespace.
However, we didn't handle that case 'correctly' anyway before.
Previously we'd also ignore the explicitly specified assembler name,
now we propagate it.  It's not clear to me what the desired semantics
would be in decorating just one of the local extern declarations this
way.  I don't think we can really do better without propagating this
aliasing property into the middle end (which is also needed for some
constexpr handling, see PR97306).  I tried that before and it turned
into a rat-hole.

PR c++/99508
gcc/cp/
* decl.c (make_rtl_for_nonlocal_decl): Propagate local-extern's
assembler name to the ns alias.
gcc/testsuite/
* g++.dg/ext/pr99508.C: New.

3 years agolibstdc++: Fix headers that can't be used as header units [PR 99413]
Jonathan Wakely [Wed, 10 Mar 2021 15:27:41 +0000 (15:27 +0000)]
libstdc++: Fix headers that can't be used as header units [PR 99413]

This adds missing includes to internal library headers which get
included from more than one other header, so that they can be compiled
as a stand-alone header unit.

For existing use cases these includes are no-ops because they're already
done by the header that includes these files. For compiling them as a
header unit, this ensures that they include what they use.

libstdc++-v3/ChangeLog:

PR libstdc++/99413
* include/bits/align.h: Include debug/assertions.h.
* include/bits/codecvt.h: Include bits/c++config.h.
* include/bits/enable_special_members.h: Likewise.
* include/bits/erase_if.h: Likewise.
* include/bits/functional_hash.h: Include <type_traits>.
* include/bits/invoke.h: Include bits/move.h.
* include/bits/ostream_insert.h: Include bits/exception_defines.h.
* include/bits/parse_numbers.h: Include <type_traits>.
* include/bits/predefined_ops.h: Include bits/c++config.h.
* include/bits/range_access.h: Include bits/stl_iterator.h.
* include/bits/stl_bvector.h: Do not include bits/stl_vector.h.
* include/bits/stl_iterator.h: Include bits/stl_iterator_base_types.h.
* include/bits/stl_uninitialized.h: Include bits/stl_algobase.h.
* include/bits/uniform_int_dist.h: Include bits/concept_check.h.
* include/bits/unique_lock.h: Include bits/std_mutex.h.
* include/debug/assertions.h: Include bits/c++config.h.

3 years agolibstdc++: Implement LWG 3530 for concept-constrained comparisons
Jonathan Wakely [Wed, 10 Mar 2021 15:27:06 +0000 (15:27 +0000)]
libstdc++: Implement LWG 3530 for concept-constrained comparisons

The proposed resolution for this library issue simplifies the
constraints for compare_three_way, ranges::equal_to, ranges::less etc.
so that they do not work with types which are convertible to pointers
but which fail to meet the usual syntactic requirements for the
comparisons.

This affects the example in PR libstdc++/93628 but doesn't fix the
problem described in that report.

libstdc++-v3/ChangeLog:

* include/bits/ranges_cmp.h (__eq_builtin_ptr_cmp): Remove.
(ranges::equal_to, ranges::not_equal_to): Do not constrain
with __eq_builtin_ptr_cmp.
(ranges::less, ranges::greater, ranges::less_equal)
(ranges::greater_equal): Do not constrain with
__less_builtin_ptr_cmp.
* libsupc++/compare (compare_three_way): Do not constrain with
__3way_builtin_ptr_cmp.
* testsuite/18_support/comparisons/object/builtin-ptr-three-way.cc: Moved to...
* testsuite/18_support/comparisons/object/lwg3530.cc: ...here.
* testsuite/20_util/function_objects/range.cmp/lwg3530.cc: New test.

3 years agoaarch64: Fix typo in aarch64.c comment
Alex Coplan [Wed, 10 Mar 2021 15:11:16 +0000 (15:11 +0000)]
aarch64: Fix typo in aarch64.c comment

This fixes a typo in the description of
aarch64_vfp_is_call_or_return_candidate.

gcc/ChangeLog:

* config/aarch64/aarch64.c (aarch64_vfp_is_call_or_return_candidate):
Fix typo in comment describing "is_ha" argument.

3 years agoAdd LRA target selector.
John David Anglin [Wed, 10 Mar 2021 14:57:01 +0000 (14:57 +0000)]
Add LRA target selector.

gcc/ChangeLog:

* doc/sourcebuild.texi: Document LRA target selector.

gcc/testsuite/ChangeLog:

PR testsuite/99292
* lib/target-supports.exp (check_effective_target_lra): New.
* gcc.c-torture/compile/asmgoto-2.c: Use LRA target selector.
* gcc.c-torture/compile/asmgoto-5.c: Likewise.
* gcc.c-torture/compile/pr98096.c: Likewise.
* gcc.dg/pr97954.c: Likewise.

3 years agoanalyzer: remove some no-longer-needed -Wno-analyzer-too-complex
David Malcolm [Wed, 10 Mar 2021 13:54:02 +0000 (08:54 -0500)]
analyzer: remove some no-longer-needed -Wno-analyzer-too-complex

A couple of analyzer testcases no longer have state explosions; updating
them accordingly in case they regress.

gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/pr94047.c: Remove "-Wno-analyzer-too-complex".
* gcc.dg/analyzer/zlib-2.c: Likewise.

3 years agodocs: add interactive vs batch distinction to UX guidelines
David Malcolm [Wed, 10 Mar 2021 13:49:29 +0000 (08:49 -0500)]
docs: add interactive vs batch distinction to UX guidelines

gcc/ChangeLog:
* doc/ux.texi: Add subsection contrasting interactive versus
batch usage of GCC.

3 years ago[testsuite] Fix target selector for pr99102.c
Joel Hutton [Wed, 10 Mar 2021 13:28:46 +0000 (13:28 +0000)]
[testsuite] Fix target selector for pr99102.c

The target selector should explicitly choose 256 bit hardware as
explicit 256 bit compiler options are used to trigger the bug.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/pr99102.c: Fix target selector.

3 years ago[Vect] Fix mask check on Scatter loads/stores
Joel Hutton [Wed, 10 Mar 2021 12:22:45 +0000 (12:22 +0000)]
[Vect] Fix mask check on Scatter loads/stores

Previously, IFN_MASK_SCATTER_STORE was used if 'loop_masks' was
non-null, but the mask used is 'final_mask'. This caused a bug where
a 'MASK_STORE' was vectorized into a 'SCATTER_STORE' instead of a
'MASK_SCATTER_STORE'. This fixes PR target/99102.

gcc/ChangeLog:

PR target/99102
* tree-vect-stmts.c (vectorizable_store): Fix scatter store mask
check condition.
(vectorizable_load): Fix gather load mask check condition.

gcc/testsuite/ChangeLog:

PR target/99102
* gcc.dg/vect/pr99102.c: New test.

3 years agotree-optimization/99510 - fix type reuse of build_aligned_type
Richard Biener [Wed, 10 Mar 2021 10:57:21 +0000 (11:57 +0100)]
tree-optimization/99510 - fix type reuse of build_aligned_type

The fix for PR94775 added more strict checking for type reuse
to check_aligned_type, specifically matching TYPE_USER_ALIGN.
But then build_aligned_type sets TYPE_USER_ALIGN on the built
variant so if the type we build an aligned variant for does not
have TYPE_USER_ALIGN we'll never re-use the newly created aligned
variant.  This results in ~35000 identical variants being created
for polyhedron doduc.

The following instead checks that the candidate has TYPE_USER_ALIGN set.

2021-03-10  Richard Biener  <rguenther@suse.de>

PR tree-optimization/99510
* tree.c (check_aligned_type): Check that the candidate
has TYPE_USER_ALIGN set instead of matching with the
original type.

3 years agoDo not assume that __float128 exists
Eric Botcazou [Wed, 10 Mar 2021 11:05:53 +0000 (12:05 +0100)]
Do not assume that __float128 exists

The code in build_round_expr implicitly assumes that __float128 exists,
which is *not* the common case among 64-bit architectures since the
"long double" type is generally already 128-bit for them.

gcc/fortran/
PR fortran/96983
* trans-intrinsic.c (build_round_expr): Do not implicitly assume
that __float128 is the 128-bit floating-point type.

3 years agoFix ICE on atomic enumeration type with LTO
Eric Botcazou [Wed, 10 Mar 2021 11:04:25 +0000 (12:04 +0100)]
Fix ICE on atomic enumeration type with LTO

This is a strange regression whereby an enumeration type declared as
atomic (or volatile) incorrectly triggers the ODR machinery for its
values in LTO mode.

gcc/ada/
* gcc-interface/decl.c (gnat_to_gnu_entity): Build a TYPE_STUB_DECL
for the main variant of an enumeration type declared as volatile.
gcc/testsuite/
* gnat.dg/specs/lto25.ads: New test.

3 years agoFix miscompilation of Ada runtime on 64-bit SPARC
Eric Botcazou [Wed, 10 Mar 2021 11:02:14 +0000 (12:02 +0100)]
Fix miscompilation of Ada runtime on 64-bit SPARC

Returning a REGMODE_NATURAL_SIZE of 4 for DFmode in 64-bit mode is
just asking for trouble because sub-word SUBREGs are always treated
differently than the others, in particular by the register allocator.

gcc/
* config/sparc/sparc.c (sparc_regmode_natural_size): Return 4 for
float and vector integer modes only if the mode is not larger.

3 years agolibstdc++: Uncomment more parts of <chrono> synopsis test
Jonathan Wakely [Wed, 10 Mar 2021 11:14:17 +0000 (11:14 +0000)]
libstdc++: Uncomment more parts of <chrono> synopsis test

libstdc++-v3/ChangeLog:

* testsuite/std/time/syn_c++20.cc: Enable synopsis checks for
C++20 calendar types.

3 years agocris: define DWARF_FRAME_REGISTERS
Hans-Peter Nilsson [Wed, 10 Mar 2021 02:54:16 +0000 (03:54 +0100)]
cris: define DWARF_FRAME_REGISTERS

When DWARF_FRAME_REGISTERS isn't defined, the default is
FIRST_PSEUDO_REGISTER which means that if you add faked
registers to the port, used for frame-context related
elimination, room is allocated for them in the register
context used for frame-unwinding, which is wasteful because
they're eliminated before the final form of the code that is
emitted.

Stopping after MOF saves two register slots in the unwind
contest, compared to the current default.  For regular C
programming this is uninteresting, but defining
DWARF_FRAME_REGISTERS now also avoids the need to remember
to define it later, when twiddling with additional faked
registers (alternatively suffering churn from comparing
differences in unwind context).  As expected, no effect on
test-results, coremark or local (C-specific)
microbenchmarks.

gcc:
* config/cris/cris.h (DWARF_FRAME_REGISTERS): Define.

3 years agoDaily bump.
GCC Administrator [Wed, 10 Mar 2021 00:16:43 +0000 (00:16 +0000)]
Daily bump.

3 years agoIRA: Process digital constraints containing more one digit
Vladimir N. Makarov [Tue, 9 Mar 2021 22:18:58 +0000 (17:18 -0500)]
IRA: Process digital constraints containing more one digit

gcc/ChangeLog:

* ira.c (ira_setup_alts, ira_get_dup_out_num): Process digital
constraints > 9.
* ira-lives.c (single_reg_class): Ditto.

3 years agoRTEMS: Fix -Werror builds
Sebastian Huber [Tue, 9 Mar 2021 18:32:26 +0000 (19:32 +0100)]
RTEMS: Fix -Werror builds

Fix build errors due to warnings such as:

gcc/config/v850/rtems.h:43: error: "RTEMS_STARTFILE_SPEC" redefined [-Werror]
   43 | #define RTEMS_STARTFILE_SPEC ""

The problem was that "gcc/config/rtems.h" was included before the
architecture-specific "gcc/config/*/rtems.h" header file on some
architectures.

gcc/

* config.gcc (aarch64-*-rtems*): Include general rtems.h after
the architecture-specific rtems.h.
(aarch64-*-rtems*): Likewise.
(arm*-*-rtems*): Likewise.
(epiphany-*-rtems*): Likewise.
(riscv*-*-rtems*): Likewise.

3 years agophiopt: Fix up conditional_replacement [PR99305]
Jakub Jelinek [Tue, 9 Mar 2021 18:13:11 +0000 (19:13 +0100)]
phiopt: Fix up conditional_replacement [PR99305]

Before my PR97690 changes, conditional_replacement would not set neg
when the nonzero arg was boolean true.
I've simplified the testing, so that it first finds the zero argument
and then checks the other argument for all the handled cases
(1, -1 and 1 << X, where the last case is what the patch added support for).
But, unfortunately I've placed the integer_all_onesp test first.
For unsigned precision 1 types such as bool integer_all_onesp, integer_onep
and integer_pow2p can all be true and the code set neg to true in that case,
which is undesirable.

The following patch tests integer_pow2p first (which is trivially true
for integer_onep too and tree_log2 in that case gives shift == 0)
and only if that isn't the case, integer_all_onesp.

2021-03-09  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/99305
* tree-ssa-phiopt.c (conditional_replacement): Test integer_pow2p
before integer_all_onesp instead of vice versa.

* g++.dg/opt/pr99305.C: New test.

3 years agors6000: Fix check_effective_target_sqrt_insn (PR99352)
Segher Boessenkool [Wed, 3 Mar 2021 20:34:32 +0000 (20:34 +0000)]
rs6000: Fix check_effective_target_sqrt_insn (PR99352)

The previous version returned true for all PowerPC.  This is incorrect.
We only support floating point square root instructions if a) we support
floating point instructions at all, and b) we have _ARCH_PPCSQ defined.

2020-03-09  Segher Boessenkool  <segher@kernel.crashing.org>

gcc/testsuite/
* lib/target-supports.exp (check_effective_target_powerpc_sqrt): New.
(check_effective_target_sqrt_insn): Use it.

3 years agoarm: fix bootstrap failure following automatic mode selection patch
Richard Earnshaw [Tue, 9 Mar 2021 17:02:53 +0000 (17:02 +0000)]
arm: fix bootstrap failure following automatic mode selection patch

Fix a signed vs unsigned comparison in last change.

gcc:
* common/config/arm/arm-common.c (arm_config_default): Change type
of 'i' to unsigned.

3 years ago[PR99454] LRA: Process separately 'g' and digital constraints > 9 in process_address_1
Vladimir N. Makarov [Tue, 9 Mar 2021 15:57:21 +0000 (10:57 -0500)]
[PR99454] LRA: Process separately 'g' and digital constraints > 9 in process_address_1

gcc/ChangeLog:

PR target/99454
* lra-constraints.c (process_address_1): Process constraint 'g'
separately and digital constraints containing more one digit.

gcc/testsuite/ChangeLog:

PR target/99454
* gcc.target/i386/pr99454.c: New.

3 years agoRe: [PATCH v2] fix Ada bootstrap on Cygwin64 (PR bootstrap/94918)
Mikael Pettersson [Tue, 9 Mar 2021 15:58:56 +0000 (08:58 -0700)]
Re: [PATCH v2] fix Ada bootstrap on Cygwin64 (PR bootstrap/94918)

gcc/ada/
* raise-gcc.c: On Cygwin include mingw32.h to prevent
windows.h from including x86intrin.h or emmintrin.h.

3 years agoc++: Fix coroutines on targetm.cxx.cdtor_return_this targets [PR99459]
Jakub Jelinek [Tue, 9 Mar 2021 15:44:27 +0000 (16:44 +0100)]
c++: Fix coroutines on targetm.cxx.cdtor_return_this targets [PR99459]

The r11-7528 build_co_await changes broke coroutines on arm*-linux-gnuabi,
2780 ^FAIL.*coroutines/ in total.
The problem is that arm is targetm.cxx.cdtor_return_this target where
both ctors and dtors in the ABI return this pointer rather than
void, and build_new_method_call_1 does:
              else if (call != error_mark_node
                       && DECL_DESTRUCTOR_P (cand->fn)
                       && !VOID_TYPE_P (TREE_TYPE (call)))
                /* An explicit call of the form "x->~X()" has type
                   "void".  However, on platforms where destructors
                   return "this" (i.e., those where
                   targetm.cxx.cdtor_returns_this is true), such calls
                   will appear to have a return value of pointer type
                   to the low-level call machinery.  We do not want to
                   change the low-level machinery, since we want to be
                   able to optimize "delete f()" on such platforms as
                   "operator delete(~X(f()))" (rather than generating
                   "t = f(), ~X(t), operator delete (t)").  */
                call = build_nop (void_type_node, call);
The new code in build_co_await relies on build_special_member_call
returned expression being a CALL_EXPR, but due to the build_nop
in there it is a NOP_EXPR around the CALL_EXPR.  It can't be stripped
with STRIP_NOPS because void has different mode from the pointer mode.

2021-03-09  Jakub Jelinek  <jakub@redhat.com>

PR c++/99459
* coroutines.cc (build_co_await): Look through NOP_EXPRs in
build_special_member_call return value to find the CALL_EXPR.
Simplify.

3 years agoFix building the RX port of gcc.
Nick Clifton [Tue, 9 Mar 2021 15:29:23 +0000 (15:29 +0000)]
Fix building the RX port of gcc.

* config/rx/rx.h (DBX_DEBUGGING_INFO): Define.
(DWARF"_DEBUGGING_INFO): Define.

3 years agoAssorted testsuite fixes
Eric Botcazou [Tue, 9 Mar 2021 15:20:53 +0000 (16:20 +0100)]
Assorted testsuite fixes

First, gcc.dg/array-quals-1.c does not pass if the compiler is configured
with --enable-default-pie because the sections change, so force -fno-pie.
Second, replace *-*-solaris* with sparc*-*-* for gfortran.dg/pr95690.f90
because this depends on the architecture rather than the OS.  Third force
SRA to trigger on Aarch64 (like PowerPC) for gnat.dg/opt39.adb.

gcc/testsuite/
* gcc.dg/array-quals-1.c: Pass -fno-pie if supported.
* gcc.dg/loop-9.c: Likewise.
* gfortran.dg/pr95690.f90: Replace *-*-solaris* with sparc*-*-*.
* gnat.dg/opt39.adb: Pass --param option for Aarch64 too.

3 years agoFix internal error on lambda function
Eric Botcazou [Tue, 9 Mar 2021 15:12:22 +0000 (16:12 +0100)]
Fix internal error on lambda function

This boils down to the RTL expander trying to take the address of a DECL
whose RTX is a register.

gcc/
PR c++/90448
* calls.c (initialize_argument_information): When the argument
is passed by reference, do not make a copy in a thunk only if
the argument is already in memory.  Remove redundant test for
the case of callee copy.

3 years ago[PR99454] LRA: Process 0..9 constraints in process_address_1
Vladimir N. Makarov [Tue, 9 Mar 2021 14:05:05 +0000 (09:05 -0500)]
[PR99454] LRA: Process 0..9 constraints in process_address_1

We need to process 0..9 constraints to fetch the right op constraint in
the function.  Also 0..9 constraints gives unknown class constraint
class which can result in skipping address normalization for memory in asm.

gcc/ChangeLog:

PR target/99454
* lra-constraints.c (process_address_1): Process 0..9 constraints
in process_address_1.

3 years agoIBM Z: arch14 fix option string used for Binutils
Andreas Krebbel [Tue, 9 Mar 2021 10:45:31 +0000 (11:45 +0100)]
IBM Z: arch14 fix option string used for Binutils

gcc/ChangeLog:

* config/s390/s390.c (struct s390_processor processor_table):
Binutils name string must not be empty.

3 years agotestsuite: Fix up pr98920.c on non-glibc or old glibc targets [PR98920]
Jakub Jelinek [Tue, 9 Mar 2021 13:14:09 +0000 (14:14 +0100)]
testsuite: Fix up pr98920.c on non-glibc or old glibc targets [PR98920]

Not all OSes have regex.h and not all OSes that do have REG_STARTEND macro support.
Conditionalize the test on that.

2021-03-09  Jakub Jelinek  <jakub@redhat.com>

PR sanitizer/98920
* c-c++-common/asan/pr98920.c: Only include regex.h if the header
exists.  If REG_STARTEND macro isn't defined, just return 0 from main
instead of the actual test.

3 years agoc++: Clarify note about -fmodules-ts [PR 99472]
Nathan Sidwell [Tue, 9 Mar 2021 13:08:09 +0000 (05:08 -0800)]
c++: Clarify note about -fmodules-ts [PR 99472]

This clarifies that c++2[03] intentionally does not enable
c++20 modules.

PR c++/99472
gcc/cp/
* parser.c (cp_parser_diagnose_invalid_type_name): Clarify
that C++20 does not yet imply modules.

3 years agoarc: Remove orphan function.
Claudiu Zissulescu [Tue, 9 Mar 2021 10:32:20 +0000 (12:32 +0200)]
arc: Remove orphan function.

Remove unused function.

gcc/
2021-03-09  Claudiu Zissulescu  <claziss@synopsys.com>

* config/arc/arc.c (arc_attr_type): Remove function.

Signed-off-by: Claudiu Zissulescu <claziss@synopsys.com>
3 years agoi386: Properly set ix86_isa_flags
Martin Liska [Mon, 8 Mar 2021 11:56:06 +0000 (12:56 +0100)]
i386: Properly set ix86_isa_flags

gcc/ChangeLog:

PR target/99464
* config/i386/i386-options.c (ix86_option_override_internal):
Set isa_flags for OPTS argument and not for the global
global_options.

gcc/testsuite/ChangeLog:

PR target/99464
* gcc.target/i386/pr99464.c: New test.

3 years agoChecked in non-final version of patch in commit 9433c844c8bcf0166567943b45576ceeeee0b131
Aaron Sawdey [Tue, 9 Mar 2021 04:18:03 +0000 (22:18 -0600)]
Checked in non-final version of patch in commit 9433c844c8bcf0166567943b45576ceeeee0b131

Not sure what I did but this corrects it to the version that I tested
and that Segher approved.

gcc/ChangeLog
* config/rs6000/predicates.md (ds_form_mem_operand): Check
in correct code.

3 years agoadd powerpc_vsx_ok requirement to undef-bool tests
Joel Brobecker [Tue, 9 Mar 2021 02:59:37 +0000 (23:59 -0300)]
add powerpc_vsx_ok requirement to undef-bool tests

These tests use -mvsx in their dg-options lists, so they are only
applicable if the -mvsx option is supported by the compiler.

for  gcc/testsuite/ChangeLog

* gcc.target/powerpc/undef-bool-2.c: Add
dg-require-effective-target powerpc_vsx_ok directive.
* g++.dg/ext/undef-bool-1.C: Add dg-require-effective-target
powerpc_vsx_ok directive.

3 years agoTighten predicates for p10 ld/cmpi fusion
Aaron Sawdey [Sun, 7 Mar 2021 20:47:31 +0000 (14:47 -0600)]
Tighten predicates for p10 ld/cmpi fusion

PR99070 is caused by a fusion pattern matching that the individual
instructions do not match when it is split later. In this case the
ld+cmpi patterns were allowing a d-form load address, which the split
condition would rightly split, however that left us with something that
could not be matched by a ds-form ld instruction, hence the ICE. This
only happened if the target cpu was not power10 -- if we were targeting
power10 then a prefixed pld instruction would get generated because that
can handle d-form. However this is not optimal code either.

So the solution is a new predicate (ds_form_mem_operand) that only
accepts what we can take as for a ds-form load. Then a small
modification of the genfusion.pl script changes the relevant
ld+cmpi patterns to use the new predicate.

gcc/ChangeLog

PR target/99070
* config/rs6000/predicates.md (ds_form_mem_operand) New
predicate.
* config/rs6000/genfusion.pl (gen_ld_cmpi_p10) Use
ds_form_mem_operand in ld/lwa patterns.
* config/rs6000/fusion.md: Regenerate file.

3 years agoruntime: cast SIGSTKSZ to uintptr
Ian Lance Taylor [Mon, 8 Mar 2021 21:58:14 +0000 (13:58 -0800)]
runtime: cast SIGSTKSZ to uintptr

In newer versions of glibc it is long, which causes a signed
comparison warning.

Fixes PR go/99458

3 years agoDaily bump.
GCC Administrator [Tue, 9 Mar 2021 00:16:57 +0000 (00:16 +0000)]
Daily bump.

3 years agoUpdate cpplib eo.po.
Joseph Myers [Mon, 8 Mar 2021 23:47:58 +0000 (23:47 +0000)]
Update cpplib eo.po.

* eo.po: Update.

3 years agoUpdate gcc de.po, sv.po.
Joseph Myers [Mon, 8 Mar 2021 23:47:01 +0000 (23:47 +0000)]
Update gcc de.po, sv.po.

* de.po, sv.po: Update.

3 years agoAdjust my email address to a personal one
Jeff Law [Mon, 8 Mar 2021 22:32:54 +0000 (15:32 -0700)]
Adjust my email address to a personal one

* MAINTAINERS: Update my email address

3 years agoPR fortran/49278 - ICE when combining DATA with default initialization
Harald Anlauf [Mon, 8 Mar 2021 20:59:20 +0000 (21:59 +0100)]
PR fortran/49278 - ICE when combining DATA with default initialization

A variable with the PARAMETER attribute may not appear in a DATA statement.

gcc/fortran/ChangeLog:

PR fortran/49278
* data.c (gfc_assign_data_value): Reject variable with PARAMETER
attribute in DATA statement.

gcc/testsuite/ChangeLog:

PR fortran/49278
* gfortran.dg/parameter_data.f90: New test.

3 years agoPR middle-end/98266 - bogus array subscript is partly outside array bounds on virtual...
Martin Sebor [Mon, 8 Mar 2021 20:37:21 +0000 (13:37 -0700)]
PR middle-end/98266 - bogus array subscript is partly outside array bounds on virtual inheritance

gcc/ChangeLog:

PR middle-end/98266
* gimple-array-bounds.cc (inbounds_vbase_memaccess_p): New function.
(array_bounds_checker::check_array_bounds): Call it.

gcc/testsuite/ChangeLog:

PR middle-end/98266
* g++.dg/warn/Warray-bounds-15.C: New test.
* g++.dg/warn/Warray-bounds-18.C: New test.
* g++.dg/warn/Warray-bounds-19.C: New test.
* g++.dg/warn/Warray-bounds-20.C: New test.
* g++.dg/warn/Warray-bounds-21.C: New test.

3 years agoPR middle-end/97631 - bogus "writing one too many bytes" warning for memcpy with...
Martin Sebor [Mon, 8 Mar 2021 20:28:52 +0000 (13:28 -0700)]
PR middle-end/97631 - bogus "writing one too many bytes" warning for memcpy with strlen argument

gcc/ChangeLog:

PR middle-end/97631
* tree-ssa-strlen.c (maybe_warn_overflow): Test rawmem.
(handle_builtin_stxncpy_strncat): Rename locals.  Determine
destination size from allocation calls.  Issue a more appropriate
kind of warning.
(handle_builtin_memcpy): Pass true as rawmem to maybe_warn_overflow.
(handle_builtin_memset): Same.

gcc/testsuite/ChangeLog:

PR middle-end/97631
* c-c++-common/Wstringop-overflow.c: Remove unexpected warnings.
Add an xfail.
* c-c++-common/Wstringop-truncation.c: Add expected warnings.
* gcc.dg/Wstringop-overflow-10.c: Also enable -Wstringop-truncation.
* gcc.dg/Wstringop-overflow-66.c: New test.
* gcc.dg/tree-ssa/strncpy-2.c: Adjust expected warning.

3 years agoc++: Add test for PR96268.
Marek Polacek [Mon, 8 Mar 2021 20:26:58 +0000 (15:26 -0500)]
c++: Add test for PR96268.

This works since the recent r11-7102, but we didn't have a test for
a template-argument context.

gcc/testsuite/ChangeLog:

PR c++/96268
* g++.dg/cpp2a/nontype-class41.C: New test.

3 years agoC++: Enable c++2b module mode [PR 99436]
Nathan Sidwell [Mon, 8 Mar 2021 19:55:26 +0000 (11:55 -0800)]
C++: Enable c++2b module mode [PR 99436]

This adds support for c++23 mode to modules, and enables such testing.

PR c++/99436
gcc/cp/
* name-lookup.c (get_cxx_dialect_name): Add cxx23.
gcc/testsuite/
* g++.dg/modules/modules.exp (MOD_STD_LIST): Add 2b.

3 years agoc++: Poor diagnostic in header-unit [PR 99468]
Nathan Sidwell [Mon, 8 Mar 2021 18:37:03 +0000 (10:37 -0800)]
c++: Poor diagnostic in header-unit [PR 99468]

We didn't specifically check for a module-decl inside a header unit.
That leads to a confusing diagostic.  Fixed thusly.

gcc/cp/
* lex.c (module_token_filter::resume): Ignore module-decls inside
header-unit.
* parser.c (cp_parser_module_declaration): Reject in header-unit.
gcc/testsuite/
* g++.dg/modules/pr99468.H: New.

3 years agors6000: Fix invalid splits when using Altivec style addresses [PR98959]
Peter Bergner [Mon, 8 Mar 2021 18:20:41 +0000 (12:20 -0600)]
rs6000: Fix invalid splits when using Altivec style addresses [PR98959]

The rs6000_emit_le_vsx_* functions assume they are not passed an Altivec
style "& ~16" address.  However, some of our expanders and splitters do
not verify we do not have an Altivec style address before calling those
functions, leading to an ICE.  The solution here is to guard the expanders
and splitters to ensure we do not call them if we're given an Altivec style
address.

2021-03-08  Peter Bergner  <bergner@linux.ibm.com>

gcc/
PR target/98959
* config/rs6000/rs6000.c (rs6000_emit_le_vsx_permute): Add an assert
to ensure we do not have an Altivec style address.
* config/rs6000/vsx.md (*vsx_le_perm_load_<mode>): Disable if passed
an Altivec style address.
(*vsx_le_perm_store_<mode>): Likewise.
(splitters after *vsx_le_perm_store_<mode>): Likewise.
(vsx_load_<mode>): Disable special expander if passed an Altivec
style address.
(vsx_store_<mode>): Likewise.

gcc/testsuite/
PR target/98959
* gcc.target/powerpc/pr98959.c: New test.

3 years agoc++: Incorrect specialization hash table [PR 99285]
Nathan Sidwell [Mon, 8 Mar 2021 18:01:21 +0000 (10:01 -0800)]
c++: Incorrect specialization hash table [PR 99285]

Class template partial specializations need to be in the
specialization hash, but not all of them.  This defers adding
streamed-in entities to the hash table, in the same way I deferred
adding the instantiation and specialization lists for 99170.

PR c++/99285
gcc/cp/
* cp-tree.h (match_mergeable_specialization)
(add_mergeable_specialization): Adjust parms.
* module.cc (trees_in::decl_value): Adjust
add_mergeable_specialization calls.
(trees_out::key_mergeable): Adjust match_mergeable_specialization
calls.
(specialization_add): Likewise.
* pt.c (match_mergeable_specialization): Do not insert.
(add_mergeable_specialization): Add to hash table here.
gcc/testsuite/
* g++.dg/modules/pr99285_a.H: New.
* g++.dg/modules/pr99285_b.H: New.

3 years agoaarch64: Fix PR99437 - tighten shift predicates for narrowing shift patterns
Kyrylo Tkachov [Mon, 8 Mar 2021 15:05:21 +0000 (15:05 +0000)]
aarch64: Fix PR99437 - tighten shift predicates for narrowing shift patterns

In this bug combine forms the (R)SHRN(2) instructions with an invalid shift amount.
The intrinsic expanders for these patterns validate the right shift amount but if the
final patterns end up being matched by combine (or other RTL passes I suppose) they
still let the wrong const_vector through.

This patch tightens up the predicates for the instructions involved by using predicates
for the right shift amount const_vectors.

gcc/ChangeLog:

PR target/99437
* config/aarch64/predicates.md (aarch64_simd_shift_imm_vec_qi): Define.
(aarch64_simd_shift_imm_vec_hi): Likewise.
(aarch64_simd_shift_imm_vec_si): Likewise.
(aarch64_simd_shift_imm_vec_di): Likewise.
* config/aarch64/aarch64-simd.md (aarch64_shrn<mode>_insn_le): Use
predicate from above.
(aarch64_shrn<mode>_insn_be): Likewise.
(aarch64_rshrn<mode>_insn_le): Likewise.
(aarch64_rshrn<mode>_insn_be): Likewise.
(aarch64_shrn2<mode>_insn_le): Likewise.
(aarch64_shrn2<mode>_insn_be): Likewise.
(aarch64_rshrn2<mode>_insn_le): Likewise.
(aarch64_rshrn2<mode>_insn_be): Likewise.

gcc/testsuite/ChangeLog:

PR target/99437
* gcc.target/aarch64/simd/pr99437.c: New test.

3 years agolibsanitizer: cherry-pick ad294e572bc5c16f9dc420cc994322de6ca3fbfb
Martin Liska [Mon, 8 Mar 2021 14:52:03 +0000 (15:52 +0100)]
libsanitizer: cherry-pick ad294e572bc5c16f9dc420cc994322de6ca3fbfb

libsanitizer/ChangeLog:

PR sanitizer/98920
* asan/asan_interceptors.cpp (COMMON_INTERCEPT_FUNCTION_VER):
Cherry pick.
(COMMON_INTERCEPT_FUNCTION_VER_UNVERSIONED_FALLBACK): Likewise.
* asan/asan_interceptors.h (ASAN_INTERCEPT_FUNC_VER_UNVERSIONED_FALLBACK): Likewise.
* sanitizer_common/sanitizer_common_interceptors.inc
(COMMON_INTERCEPT_FUNCTION_GLIBC_VER_MIN): Likewise.
(INIT_REGEX): Likewise.
* tsan/tsan_interceptors_posix.cpp (COMMON_INTERCEPT_FUNCTION_VER_UNVERSIONED_FALLBACK):
Likewise.

gcc/testsuite/ChangeLog:

PR sanitizer/98920
* c-c++-common/asan/pr98920.c: New test.

3 years ago[PR99422] LRA: Skip modifiers when processing memory address.
Vladimir N. Makarov [Mon, 8 Mar 2021 14:24:57 +0000 (09:24 -0500)]
[PR99422] LRA: Skip modifiers when processing memory address.

  Function process_address_1 can wrongly look at constraint modifiers
instead of the 1st constraint itself.  The patch solves the problem.

gcc/ChangeLog:

PR target/99422
* lra-constraints.c (skip_contraint_modifiers): New function.
(process_address_1): Use it before lookup_constraint call.

3 years agoi386: Enable UINTR and HRESET for -march that supports it
Martin Liska [Mon, 8 Mar 2021 11:08:37 +0000 (12:08 +0100)]
i386: Enable UINTR and HRESET for -march that supports it

gcc/ChangeLog:

PR target/99463
* config/i386/i386-options.c (ix86_option_override_internal):
Enable UINTR and HRESET for -march that supports it.

gcc/testsuite/ChangeLog:

PR target/99463
* gcc.target/i386/pr99463-2.c: New test.
* gcc.target/i386/pr99463.c: New test.

3 years agoIBM Z: Fix usage of "f" constraint with long doubles
Ilya Leoshkevich [Thu, 10 Dec 2020 14:43:59 +0000 (15:43 +0100)]
IBM Z: Fix usage of "f" constraint with long doubles

After switching the s390 backend to store long doubles in vector
registers, "f" constraint broke when used with the former: long doubles
correspond to TFmode, which in combination with "f" corresponds to
hard regs %v0-%v15, however, asm users expect a %f0-%f15 pair.

Fix by using TARGET_MD_ASM_ADJUST hook to convert TFmode values to
FPRX2mode and back.

gcc/ChangeLog:

2020-12-14  Ilya Leoshkevich  <iii@linux.ibm.com>

* config/s390/s390.c (f_constraint_p): New function.
(s390_md_asm_adjust): Implement TARGET_MD_ASM_ADJUST.
(TARGET_MD_ASM_ADJUST): Likewise.

gcc/testsuite/ChangeLog:

2020-12-14  Ilya Leoshkevich  <iii@linux.ibm.com>

* gcc.target/s390/vector/long-double-asm-commutative.c: New
test.
* gcc.target/s390/vector/long-double-asm-earlyclobber.c: New
test.
* gcc.target/s390/vector/long-double-asm-in-out.c: New test.
* gcc.target/s390/vector/long-double-asm-inout.c: New test.
* gcc.target/s390/vector/long-double-asm-matching.c: New test.
* gcc.target/s390/vector/long-double-asm-regmem.c: New test.
* gcc.target/s390/vector/long-double-volatile-from-i64.c: New
test.

3 years agotree-nested: Update assert for Fortran module vars [PR97927]
Tobias Burnus [Mon, 8 Mar 2021 12:05:48 +0000 (13:05 +0100)]
tree-nested: Update assert for Fortran module vars [PR97927]

gcc/ChangeLog:

PR fortran/97927
* tree-nested.c (convert_local_reference_stmt): Avoid calling
lookup_field_for_decl for Fortran module (= namespace context).

gcc/testsuite/ChangeLog:

PR fortran/97927
* gfortran.dg/module_variable_3.f90: New test.

3 years agoIBM Z: Fix vcond-shift.c testcase.
Andreas Krebbel [Mon, 8 Mar 2021 11:49:22 +0000 (12:49 +0100)]
IBM Z: Fix vcond-shift.c testcase.

Due to a common code change the comparison in the testcase is emitted
via vec_cmp instead of vcond.  The testcase checks for an optimization
currently only available via vcond.

Fixed by implementing the same optimization also in
s390_expand_vec_compare.

gcc/ChangeLog:

* config/s390/s390.c (s390_expand_vec_compare): Implement <0
comparison with arithmetic right shift.
(s390_expand_vcond): No need for a force_reg anymore.
s390_vec_compare will do it.
* config/s390/vector.md ("vec_cmp<mode><tointvec>"): Accept also
immediate operands.

3 years agoDaily bump.
GCC Administrator [Mon, 8 Mar 2021 00:16:34 +0000 (00:16 +0000)]
Daily bump.

3 years agoi386: Fix some -mavx512vl -mno-avx512bw bugs [PR99321]
Jakub Jelinek [Sun, 7 Mar 2021 09:27:28 +0000 (10:27 +0100)]
i386: Fix some -mavx512vl -mno-avx512bw bugs [PR99321]

As I wrote in the mail with the previous PR99321 fix, we have various
bugs where we emit instructions that need avx512bw and avx512vl
ISAs when compiling with -mavx512vl -mno-avx512bw.

Without the following patch, the attached testcase fails with:
/tmp/ccW4PsfG.s: Assembler messages:
/tmp/ccW4PsfG.s:9: Error: unsupported instruction `vpaddb'
/tmp/ccW4PsfG.s:20: Error: unsupported instruction `vpaddb'
/tmp/ccW4PsfG.s:31: Error: unsupported instruction `vpaddw'
/tmp/ccW4PsfG.s:42: Error: unsupported instruction `vpaddw'
/tmp/ccW4PsfG.s:53: Error: unsupported instruction `vpsubb'
/tmp/ccW4PsfG.s:64: Error: unsupported instruction `vpsubb'
/tmp/ccW4PsfG.s:75: Error: unsupported instruction `vpsubw'
/tmp/ccW4PsfG.s:86: Error: unsupported instruction `vpsubw'
/tmp/ccW4PsfG.s:97: Error: unsupported instruction `vpmullw'
/tmp/ccW4PsfG.s:108: Error: unsupported instruction `vpmullw'
/tmp/ccW4PsfG.s:133: Error: unsupported instruction `vpminub'
/tmp/ccW4PsfG.s:144: Error: unsupported instruction `vpminuw'
/tmp/ccW4PsfG.s:155: Error: unsupported instruction `vpminuw'
/tmp/ccW4PsfG.s:166: Error: unsupported instruction `vpminsb'
/tmp/ccW4PsfG.s:177: Error: unsupported instruction `vpminsb'
/tmp/ccW4PsfG.s:202: Error: unsupported instruction `vpminsw'
/tmp/ccW4PsfG.s:227: Error: unsupported instruction `vpmaxub'
/tmp/ccW4PsfG.s:238: Error: unsupported instruction `vpmaxuw'
/tmp/ccW4PsfG.s:249: Error: unsupported instruction `vpmaxuw'
/tmp/ccW4PsfG.s:260: Error: unsupported instruction `vpmaxsb'
/tmp/ccW4PsfG.s:271: Error: unsupported instruction `vpmaxsb'
/tmp/ccW4PsfG.s:296: Error: unsupported instruction `vpmaxsw'

We already have Yw constraint which is equivalent to v for
-mavx512bw -mavx512vl and to nothing otherwise, per discussions
this patch changes it to stand for x otherwise.  As it is an
undocumented internal constraint, hopefully it won't affect
any inline asm in the wild.
For the instructions that need both we need to use Yw and
v for modes that don't need that.

2021-03-07  Jakub Jelinek  <jakub@redhat.com>

PR target/99321
* config/i386/constraints.md (Yw): Use SSE_REGS if TARGET_SSE
but TARGET_AVX512BW or TARGET_AVX512VL is not set.  Adjust description
and comment.
* config/i386/sse.md (v_Yw): New define_mode_attr.
(*<insn><mode>3, *mul<mode>3<mask_name>, *avx2_<code><mode>3,
*sse4_1_<code><mode>3<mask_name>): Use <v_Yw> instead of v
in constraints.
* config/i386/mmx.md (mmx_pshufw_1, *vec_dupv4hi): Use Yw instead of
xYw in constraints.

* lib/target-supports.exp
(check_effective_target_assembler_march_noavx512bw): New effective
target.
* gcc.target/i386/avx512vl-pr99321-1.c: New test.

3 years agoDaily bump.
GCC Administrator [Sun, 7 Mar 2021 00:16:24 +0000 (00:16 +0000)]
Daily bump.

3 years agoc++: Fix constexpr evaluation of pre-increment when !lval [PR99287]
Patrick Palka [Sat, 6 Mar 2021 22:09:07 +0000 (17:09 -0500)]
c++: Fix constexpr evaluation of pre-increment when !lval [PR99287]

Here, during cxx_eval_increment_expression (with lval=false) of
++__first where __first is &"mystr"[0], we correctly update __first
to &"mystr"[1] but we end up returning &"mystr"[0] + 1 instead of
&"mystr"[1].  This unreduced return value inhibits other pointer
arithmetic folding during later constexpr evaluation, which ultimately
causes the constexpr evaluation to fail.

It turns out the simplification of &"mystr"[0] + 1 to &"mystr"[1]
is performed by cxx_fold_pointer_plus_expression, not by fold_build2.
So we perform this simplification during constexpr evaluation of
the temporary MODIFY_EXPR (during which we assign to __first the
simplified value), but then we return 'mod' which has only been folded
via fold_build2 and hasn't gone through cxx_fold_pointer_plus_expression.

This patch fixes this by updating 'mod' with the result of the
MODIFY_EXPR evaluation appropriately, so that it captures any additional
folding of the expression when !lval.  We now need to be wary of this
evaluation failing and returning e.g. the MODIFY_EXPR or NULL_TREE; it
seems checking *non_constant_p should cover our bases here and is
generally prudent.

gcc/cp/ChangeLog:

PR c++/99287
* constexpr.c (cxx_eval_increment_expression): Pass lval when
evaluating the MODIFY_EXPR, and update 'mod' with the result of
this evaluation.  Check *non_constant_p afterwards.  For prefix
ops, just return 'mod'.

gcc/testsuite/ChangeLog:

PR c++/99287
* g++.dg/cpp2a/constexpr-99287.C: New test.

Co-authored-by: Jakub Jelinek <jakub@redhat.com>
3 years agomiddle-end: Pretty-print address space of aggregates
Julian Brown [Sat, 6 Mar 2021 19:07:26 +0000 (19:07 +0000)]
middle-end: Pretty-print address space of aggregates

This patch adds printing of "<address-space-N>" markers for aggregates
in non-default address spaces.

gcc/
* tree-pretty-print.c (dump_generic_node): Emit non-generic
address space info for aggregates.

3 years agod: Don't set default flag_complex_method.
Iain Buclaw [Sat, 6 Mar 2021 17:18:44 +0000 (18:18 +0100)]
d: Don't set default flag_complex_method.

D doesn't need C99-like requirements for complex multiply and divide,
the default set by common.opt is sufficient enough.

gcc/d/ChangeLog:

* d-lang.cc (d_init_options_struct): Don't set default
flag_complex_method.

3 years agolibgcov: Fix build on Darwin [PR99406]
Jakub Jelinek [Sat, 6 Mar 2021 15:22:27 +0000 (16:22 +0100)]
libgcov: Fix build on Darwin [PR99406]

As reported, bootstrap currently fails on older Darwin because MAP_ANONYMOUS
is not defined.

The following is what gcc/system.h does, so I think it should work for
libgcov.

2021-03-06  Jakub Jelinek  <jakub@redhat.com>

PR gcov-profile/99406
* libgcov.h (MAP_FAILED, MAP_ANONYMOUS): If HAVE_SYS_MMAN_H is
defined, define these macros if not defined already.

3 years agolibstdc++: Improve std::rot[lr] [PR99396]
Jakub Jelinek [Sat, 6 Mar 2021 10:11:30 +0000 (11:11 +0100)]
libstdc++: Improve std::rot[lr] [PR99396]

As can be seen on:

unsigned char f1 (unsigned char x, int y) { return std::rotl (x, y); }
unsigned char f2 (unsigned char x, int y) { return std::rotr (x, y); }
unsigned short f3 (unsigned short x, int y) { return std::rotl (x, y); }
unsigned short f4 (unsigned short x, int y) { return std::rotr (x, y); }
unsigned int f5 (unsigned int x, int y) { return std::rotl (x, y); }
unsigned int f6 (unsigned int x, int y) { return std::rotr (x, y); }
unsigned long int f7 (unsigned long int x, int y) { return std::rotl (x, y); }
unsigned long int f8 (unsigned long int x, int y) { return std::rotr (x, y); }
unsigned long long int f9 (unsigned long long int x, int y) { return std::rotl (x, y); }
unsigned long long int f10 (unsigned long long int x, int y) { return std::rotr (x, y); }
//unsigned __int128 f11 (unsigned __int128 x, int y) { return std::rotl (x, y); }
//unsigned __int128 f12 (unsigned __int128 x, int y) { return std::rotr (x, y); }

constexpr auto a = std::rotl (1234U, 0);
constexpr auto b = std::rotl (1234U, 5);
constexpr auto c = std::rotl (1234U, -5);
constexpr auto d = std::rotl (1234U, -__INT_MAX__ - 1);
the current <bit> definitions of std::__rot[lr] aren't pattern recognized
as rotates, they are too long/complex for that, starting with signed modulo,
special case for 0 and different cases for positive and negative.

For types with power of two bits the following patch adds definitions that
the compiler can pattern recognize and turn e.g. on x86_64 into ro[lr][bwlq]
instructions.  For weirdo types like unsigned __int20 etc. it keeps the
current definitions.

2021-03-06  Jakub Jelinek  <jakub@redhat.com>

PR libstdc++/99396
* include/std/bit (__rotl, __rotr): Add optimized variants for power of
two _Nd which the compiler can pattern match the rotates.

3 years agoc++: Fix tsubsting member variable template-id [PR96330]
Patrick Palka [Sat, 6 Mar 2021 05:07:43 +0000 (00:07 -0500)]
c++: Fix tsubsting member variable template-id [PR96330]

This makes tsubst_copy appropriately handle a variable template-id, which
in turn fixes tsubsting a COMPONENT_REF whose member operand is known at
parse time to be a variable template-id, as in the initialization of 'x'
in the first testcase.  Previously, we rejected this testcase with the
error "foo_t::bar<T> is not a function template", issued from
lookup_template_fuction.

We were already properly handling the analagous case where the object
operand of the COMPONENT_REF is dependent (and so the member operand is
a dependent template name), but there doesn't seems to be existing test
coverage for this, hence the second testcase below.

gcc/cp/ChangeLog:

PR c++/96330
* pt.c (tsubst_copy) <case TEMPLATE_ID_EXPR>: Rename local
variable 'fn' to 'tmpl'.  Handle a variable template-id by
calling lookup_template_variable.

gcc/testsuite/ChangeLog:

PR c++/96330
* g++.dg/cpp1y/var-templ68.C: New test.
* g++.dg/cpp1y/var-templ68a.C: New test.

Co-authored-by: Jakub Jelinek <jakub@redhat.com>
3 years agoc++: adc_unify deduction with constrained auto [PR99365]
Patrick Palka [Sat, 6 Mar 2021 05:07:35 +0000 (00:07 -0500)]
c++: adc_unify deduction with constrained auto [PR99365]

My recent r11-7454 changed the way do_auto_deduction handles constrained
placeholders during template argument deduction (context == adc_unify)
when processing_template_decl != 0.  Before the patch, we would just
ignore the constraints on the placeholder, and return the deduced type.
After the patch, we now punt and return the original placeholder type

While this change fixed instances where we'd prematurely resolve a
constrained placeholder return or variable type with non-dependent
initializer at template parse time (such as PR96444), it broke the
adc_unify callers that rely on the previous behavior.

This patch restores the previous behavior during adc_unify deduction
while retaining the new behavior only during adc_variable_type or
adc_return_type deduction.

We additionally now need to pass the outer template arguments to
do_auto_deduction during unify, for sake of constraint checking.
But we want to avoid substituting these outer arguments into type
when the caller has already done so, so this patch adds a
TEMPLATE_TYPE_LEVEL check to do_auto_deduction to that effect.

This above is enough to fix partial specialization of non-nested
templates with constrained 'auto' template parameters, but it doesn't
fix the nested template case, ultimately because
most_specialized_partial_spec passes only the innermost template
arguments to get_partial_spec_bindings, and so outer_targs during
do_auto_deduction (called from unify) contains only the innermost
template arguments, and this breaks satisfaction.  Fixing this properly
is perhaps too risky at this stage, so this patch adds a hack to
do_auto_deduction to compensate for callers that don't supply all outer
template arguments.  The goal of this hack is to ensure placeholder type
constraint checking continues to work whenever it worked before
r11-7454, namely whenever the constraint is non-dependent.

Finally, this patch allows do_auto_deduction to resolve a constrained
placeholder type ahead of time (at template parse time), as long as the
constraint is non-dependent.

gcc/cp/ChangeLog:

PR c++/99365
* pt.c (unify) <case TEMPLATE_TYPE_PARM>: Pass targs as
outer_targs to do_auto_deduction.
(placeholder_type_constraint_dependent_p): Define.
(do_auto_deduction): When processing_template_decl != 0
and context is adc_unify and we have constraints, pretend the
constraints are satisfied instead of punting.  Otherwise don't
punt unless placeholder_type_constraint_dependent_p holds.
Add some clarifying sanity checks.  Add a hack to add missing
outermost template levels to outer_args before checking
satisfaction.  Don't substitute outer_targs into type if it's
already been done.

gcc/testsuite/ChangeLog:

PR c++/99365
* g++.dg/cpp2a/concepts-partial-spec9.C: New test.
* g++.dg/cpp2a/concepts-placeholder4.C: New test.

3 years agocris: don't define MAX_FIXED_MODE_SIZE
Hans-Peter Nilsson [Thu, 4 Mar 2021 21:58:42 +0000 (22:58 +0100)]
cris: don't define MAX_FIXED_MODE_SIZE

It's been 32 ever since the CRIS port was committed.
A TODO-item of mine has been to check whether the
non-default setting of MAX_FIXED_MODE_SIZE makes sense
wrt. performance and/or code-size with a modern gcc.  It
doesn't, so it goes.  The setting is now the default,
GET_MODE_BITSIZE (DImode) (defaults.h) i.e. 64.

Measurements at r11-7500 (f3641ac70eb0) on coremark with
"-O2 -march=v10 -mno-mul-bug-workaround" shows 0.04%
performance improvement with this change, and by inspection
the effect is that unused and/or unneeded stack-frames are
eliminated more often in the floating-point library (not in
the coremark main loop, thus the marginal improvement).  The
floating-point library is full of 64-bit unions used to pick
apart floating point numbers, so this kind of makes sense.

Inspection of a simulator trace shows that this is indeed
the only effect in coremark.  Other local micro-benchmarks
agree as to the net effect (no traces were inspected
though), and the most floating-point-heavy test shows an 8%
improvement.  These effects are of course subject to gcc
core tweaks and may make sense to be adjusted again in a
future release.

While MAX_FIXED_MODE_SIZE is IMO supposed to be an optional
macro for performance, setting it to anything smaller than
twice the size of an address exposes bad decisions in gcc
middle end, sometimes leading to internal compiler errors.
(It being set to 32 should *not* affect use of DImode as an
integer mode; it's for "integer machine modes of this size
or smaller can be used for structures and unions with the
appropriate sizes".)  Thus, with the default 64 instead of
32, there are two tests that now pass for the first time:
gcc.dg/attr-vector_size.c and gcc.dg/tree-ssa/pr93121-1.c.

gcc:
* config/cris/cris.h (MAX_FIXED_MODE_SIZE): Don't define.

3 years agogcc.target/cris/pr93372-1.c: Adjust expectations for eliminated stack-frame
Hans-Peter Nilsson [Fri, 5 Mar 2021 22:01:07 +0000 (23:01 +0100)]
gcc.target/cris/pr93372-1.c: Adjust expectations for eliminated stack-frame

See comment.

* gcc.target/cris/pr93372-1.c: Adjust expected assembler result
to allow an eliminated stack-frame.

3 years agoDaily bump.
GCC Administrator [Sat, 6 Mar 2021 00:16:43 +0000 (00:16 +0000)]
Daily bump.

3 years agoc++: Pointer-to-member fn conversion with noexcept [PR99374]
Marek Polacek [Thu, 4 Mar 2021 19:25:01 +0000 (14:25 -0500)]
c++: Pointer-to-member fn conversion with noexcept [PR99374]

The issue in this PR is that we wrongly reject converting pointers to
member function of incomplete types, one of which has noexcept.  Recall
that pointers (including pointers to member functions) to non-throwing
functions can be implicitly converted to potentially-throwing functions
(but not vice versa).

We reject the conversion when called from can_convert_arg_bad because
standard_conversion can't create such a conversion.  It comes down to
the DERIVED_FROM_P check in the TYPE_PTRMEMFUNC_P block.  It considers
every class derived from itself, but not when the class is incomplete.
But surely we want to reach fnptr_conv_p when tbase is fbase (one of
them could be an alias to the other so use same_type_p instead of ==).

Another approach would be to not perform DERIVED_FROM_P at all when
either tbase or fbase are incomplete (so perhaps something like at the
end of ptr_reasonably_similar).

gcc/cp/ChangeLog:

PR c++/99374
* call.c (standard_conversion): When converting pointers to
member, don't return NULL when the bases are equivalent but
incomplete.

gcc/testsuite/ChangeLog:

PR c++/99374
* g++.dg/cpp1z/noexcept-type23.C: New test.

3 years agoc++: ICE with -Wshadow and enumerator in template [PR99120]
Marek Polacek [Fri, 5 Mar 2021 15:41:41 +0000 (10:41 -0500)]
c++: ICE with -Wshadow and enumerator in template [PR99120]

We crash here, because in a template, an enumerator doesn't have
a type until we've called finish_enum_value_list.  But our -Wshadow
implementation, check_local_shadow, is called when we pushdecl in
build_enumerator, which takes place before finish_enum_value_list.

gcc/cp/ChangeLog:

PR c++/99120
* name-lookup.c (check_local_shadow): Check if the type of decl
is non-null before checking TYPE_PTR*.

gcc/testsuite/ChangeLog:

PR c++/99120
* g++.dg/warn/Wshadow-17.C: New test.

3 years agotestsuite: Update testcase for PR96078 fix [PR99363]
Jason Merrill [Fri, 5 Mar 2021 22:07:25 +0000 (17:07 -0500)]
testsuite: Update testcase for PR96078 fix [PR99363]

My fix for PR96078 made us stop warning about flatten on an alias if the
target has the alias, which is exactly the case tested here.  So let's
remove the expected warning and add a similar case which does warn.

gcc/testsuite/ChangeLog:

PR c/99363
* gcc.dg/attr-flatten-1.c: Adjust.

3 years agoopenmp: Avoid ICEs due to orphaned labels in OpenMP regions [PR99322]
Jakub Jelinek [Fri, 5 Mar 2021 20:52:35 +0000 (21:52 +0100)]
openmp: Avoid ICEs due to orphaned labels in OpenMP regions [PR99322]

When performing cfg cleanup at the end of cfg pass, if there are any OpenMP
regions and some basic blocks are unreachable and contain forced labels,
remove_bb moves the labels to previous bb, but if the two bb belong to different
OpenMP regions, that means it will end up in a different function from where
it was assumed to be and checked e.g. during gimplification or OpenMP region
SESE checking.

The following patch will place the labels to some bb from the right OpenMP
region if the previous bb is not that.  I think it should happen very rarely,
normally the bbs from each OpenMP region should be from the before-cfg pass
adjacent and the problems will usually be only if the OpenMP regions are
no-return, so I hope it isn't fatal that it searches through all bbs on the miss.
If it turns out to be a problem, it can always lazily create some better data
structure and maintain it through bb removals when it reaches that case the
first time.

2021-03-05  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/99322
* tree-cfg.c (bb_to_omp_idx): New variable.
(execute_build_cfg): Release the bb_to_omp_idx vector after
cleanup_tree_cfg returns.
(handle_abnormal_edges): Remove bb_to_omp_idx argument, adjust
for bb_to_omp_idx being a vec<int> instead of pointer to array
of ints.
(make_edges): Remove bb_to_omp_idx local variable, don't pass
it to handle_abnormal_edges, adjust for bb_to_omp_idx being a
vec<int> instead of pointer to array of ints and don't free/release
it at the end.
(remove_bb): When removing a bb and placing forced label somewhere
else, ensure it is put into the same OpenMP region during cfg
pass if possible or to entry successor as fallback.  Unregister
bb from bb_to_omp_idx.

* c-c++-common/gomp/pr99322.c: New test.

3 years agoc++: Duplicate namespace bindings [PR 99245]
Nathan Sidwell [Fri, 5 Mar 2021 20:49:34 +0000 (12:49 -0800)]
c++: Duplicate namespace bindings [PR 99245]

Header units can declare the same entity, and this can lead to one of
them containing a (non-using) binding to an import.  If one gets the
cluster ordering just right, an assert will trigger.  Relax that assert.

PR c++/99245
gcc/cp/
* module.cc (module_state::write_cluster): Relax binding assert.
gcc/testsuite/
* g++.dg/modules/pr99245_a.H: New.
* g++.dg/modules/pr99245_b.H: New.

3 years agoPR libfortran/99218 - matmul on temporary array accesses invalid memory
Harald Anlauf [Fri, 5 Mar 2021 19:57:54 +0000 (20:57 +0100)]
PR libfortran/99218 - matmul on temporary array accesses invalid memory

Do not invoke tuned rank-2 times rank-2 matmul if rank(b) == 1.

libgfortran/ChangeLog:

PR libfortran/99218
* m4/matmul_internal.m4: Invoke tuned matmul only for rank(b)>1.
* generated/matmul_c10.c: Regenerated.
* generated/matmul_c16.c: Likewise.
* generated/matmul_c4.c: Likewise.
* generated/matmul_c8.c: Likewise.
* generated/matmul_i1.c: Likewise.
* generated/matmul_i16.c: Likewise.
* generated/matmul_i2.c: Likewise.
* generated/matmul_i4.c: Likewise.
* generated/matmul_i8.c: Likewise.
* generated/matmul_r10.c: Likewise.
* generated/matmul_r16.c: Likewise.
* generated/matmul_r4.c: Likewise.
* generated/matmul_r8.c: Likewise.
* generated/matmulavx128_c10.c: Likewise.
* generated/matmulavx128_c16.c: Likewise.
* generated/matmulavx128_c4.c: Likewise.
* generated/matmulavx128_c8.c: Likewise.
* generated/matmulavx128_i1.c: Likewise.
* generated/matmulavx128_i16.c: Likewise.
* generated/matmulavx128_i2.c: Likewise.
* generated/matmulavx128_i4.c: Likewise.
* generated/matmulavx128_i8.c: Likewise.
* generated/matmulavx128_r10.c: Likewise.
* generated/matmulavx128_r16.c: Likewise.
* generated/matmulavx128_r4.c: Likewise.
* generated/matmulavx128_r8.c: Likewise.

gcc/testsuite/ChangeLog:

PR libfortran/99218
* gfortran.dg/matmul_21.f90: New test.

3 years agoc++: Local instantiations of imported specializations [PR 99377]
Nathan Sidwell [Fri, 5 Mar 2021 18:34:23 +0000 (10:34 -0800)]
c++: Local instantiations of imported specializations [PR 99377]

This turned out to be the function version of the previous fix.  We
can import an implicit specialization declaration that we need to
instantiate.  We must mark the instantiation so we remember to stream
it.

PR c++/99377
gcc/cp/
* pt.c (instantiate_decl): Call set_instantiating_module.
gcc/testsuite/
* g++.dg/modules/pr99377_a.H: New.
* g++.dg/modules/pr99377_b.C: New.
* g++.dg/modules/pr99377_c.C: New.

3 years agocoroutines : Adjust constraints on when to build ctors [PR98118].
Iain Sandoe [Wed, 17 Feb 2021 15:13:57 +0000 (15:13 +0000)]
coroutines : Adjust constraints on when to build ctors [PR98118].

PR98118 shows that TYPE_NEEDS_CONSTRUCTING is necessary but not
sufficient.  Use type_build_ctor_call() instead.

gcc/cp/ChangeLog:

PR c++/98118
* coroutines.cc (build_co_await): Use type_build_ctor_call()
to determine cases when a CTOR needs to be built.
(flatten_await_stmt): Likewise.
(morph_fn_to_coro): Likewise.

gcc/testsuite/ChangeLog:

PR c++/98118
* g++.dg/coroutines/pr98118.C: New test.

3 years agocoroutines : Do not accept throwing final await expressions [PR95616].
Iain Sandoe [Mon, 15 Feb 2021 17:11:31 +0000 (17:11 +0000)]
coroutines : Do not accept throwing final await expressions [PR95616].

From the PR:

The wording of [dcl.fct.def.coroutine]/15 states:
 * The expression co_await promise.final_suspend() shall not be
   potentially-throwing ([except.spec]).

See http://eel.is/c++draft/dcl.fct.def.coroutine#15
and http://eel.is/c++draft/except.spec#6

ie. all of the following must be declared noexcept (if they form part of the await-expression):
- promise_type::final_suspend()
- finalSuspendObj.operator co_await()
- finalSuspendAwaiter.await_ready()
- finalSuspendAwaiter.await_suspend()
- finalSuspendAwaiter.await_resume()
- finalSuspedObj destructor
- finalSuspendAwaiter destructor

This implements the checks for these cases and rejects such code with
a diagnostic if exceptions are enabled.

gcc/cp/ChangeLog:

PR c++/95616
* coroutines.cc (coro_diagnose_throwing_fn): New helper.
(coro_diagnose_throwing_final_aw_expr): New helper.
(build_co_await): Diagnose throwing final await expression
components.
(build_init_or_final_await): Diagnose a throwing promise
final_suspend() call.

gcc/testsuite/ChangeLog:

PR c++/95616
* g++.dg/coroutines/pr95616-0-no-exceptions.C: New test.
* g++.dg/coroutines/pr95616-0.C: New test.
* g++.dg/coroutines/pr95616-1-no-exceptions.C: New test.
* g++.dg/coroutines/pr95616-1.C: New test.
* g++.dg/coroutines/pr95616-2.C: New test.
* g++.dg/coroutines/pr95616-3-no-exceptions.C: New test.
* g++.dg/coroutines/pr95616-3.C: New test.
* g++.dg/coroutines/pr95616-4.C: New test.
* g++.dg/coroutines/pr95616-5.C: New test.
* g++.dg/coroutines/pr95616-6.C: New test.

3 years agocoroutines : Handle exceptions throw before the first await_resume() [PR95615].
Iain Sandoe [Mon, 15 Feb 2021 15:09:27 +0000 (15:09 +0000)]
coroutines : Handle exceptions throw before the first await_resume() [PR95615].

The coroutine body is wrapped in a try-catch block which is responsible for
handling any exceptions thrown by the original function body.  Originally, the
initial suspend expression was outside this, but an amendement to the standard
places the await_resume call inside and eveything else outside.

This means that any exception thrown prior to the initial suspend expression
await_resume() will propagate to the ramp function.  However, some portion of
the coroutine state will exist at that point (how much depends on where the
exception is thrown from).  For example, we might have some frame parameter
copies, or the promise object or the return object any of which might have a
non-trivial DTOR.  Also the frame itself needs to be deallocated. This patch
fixes the handling of these cases.

gcc/cp/ChangeLog:

PR c++/95615
* coroutines.cc (struct param_info): Track parameter copies that need
a DTOR.
(coro_get_frame_dtor): New helper function factored from build_actor().
(build_actor_fn): Use coro_get_frame_dtor().
(morph_fn_to_coro): Track parameters that need DTORs on exception,
likewise the frame promise and the return object.  On exception, run the
DTORs for these, destroy the frame and then rethrow the exception.

gcc/testsuite/ChangeLog:

PR c++/95615
* g++.dg/coroutines/torture/pr95615-01.C: New test.
* g++.dg/coroutines/torture/pr95615-02.C: New test.
* g++.dg/coroutines/torture/pr95615-03.C: New test.
* g++.dg/coroutines/torture/pr95615-04.C: New test.
* g++.dg/coroutines/torture/pr95615-05.C: New test.

3 years ago[PR99378] LRA: Skip decomposing address for asm insn operand with unknown constraint.
Vladimir N. Makarov [Fri, 5 Mar 2021 16:41:25 +0000 (11:41 -0500)]
[PR99378] LRA: Skip decomposing address for asm insn operand with unknown constraint.

  Function get_constraint_type returns CT__UNKNOWN for empty constraint
and CT_FIXED_FORM for "X".  So process_address_1 skipped
decompose_mem_address only for "X" constraint.  To do the same for empty
constraint, skip decompose_mem_address for CT__UNKNOWN.

gcc/ChangeLog:

PR target/99378
* lra-constraints.c (process_address_1): Skip decomposing address
for asm insn operand with unknown constraint.

gcc/testsuite/ChangeLog:

PR target/99378
* gcc.target/i386/pr99123-2.c: New.