platform/upstream/gcc.git
3 years agoDo not assume loop header threading in backward threader.
Aldy Hernandez [Fri, 3 Sep 2021 10:07:49 +0000 (12:07 +0200)]
Do not assume loop header threading in backward threader.

The registry's thread_through_all_blocks() has a may_peel_loop_headers
argument.  When refactoring the backward threader code, I removed this
argument for the local passthru method because it was always TRUE.  This
may not necessarily be true in the future, if the backward threader is
called from another context.  This patch removes the default definition,
in favor of an argument that is exactly the same as the identically
named function in tree-ssa-threadupdate.c.  I think this also makes it
less confusing when looking at both methods across the source base.

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-threadbackward.c (back_threader::thread_through_all_blocks):
Add may_peel_loop_headers.
(back_threader_registry::thread_through_all_blocks): Same.
(try_thread_blocks): Pass may_peel_loop_headers argument.
(pass_early_thread_jumps::execute): Same.

3 years agoAbstract PHI and forwarder block checks in jump threader.
Aldy Hernandez [Fri, 3 Sep 2021 09:55:11 +0000 (11:55 +0200)]
Abstract PHI and forwarder block checks in jump threader.

This patch abstracts out a couple common idioms in the forward
threader that I found useful while navigating the code base.

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-threadedge.c (has_phis_p): New.
(forwarder_block_p): New.
(potentially_threadable_block): Call forwarder_block_p.
(jump_threader::thread_around_empty_blocks): Call has_phis_p.
(jump_threader::thread_through_normal_block): Call
forwarder_block_p.

3 years agoImprove backwards threader debugging dumps.
Aldy Hernandez [Fri, 3 Sep 2021 08:57:33 +0000 (10:57 +0200)]
Improve backwards threader debugging dumps.

This patch adds debugging helpers to the backwards threader.  I have
also noticed that profitable_path_p() can bail early on paths that
crosses loops and leave the dump of blocks incomplete.  Fixed as
well.

Unfortunately the new methods cannot be marked const, because we call
the solver's dump which is not const.  I believe this was because the
ranger dump calls m_cache.block_range().  This could probably use a
cleanup at a later time.

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-threadbackward.c (back_threader::dump): New.
(back_threader::debug): New.
(back_threader_profitability::profitable_path_p): Dump blocks
even if we are bailing early.

3 years agoDump reason why threads are being cancelled and abstract code.
Aldy Hernandez [Fri, 3 Sep 2021 08:09:41 +0000 (10:09 +0200)]
Dump reason why threads are being cancelled and abstract code.

We are inconsistent on dumping out reasons why a thread was canceled.
This makes debugging jump threading problems harder because paths can be
canceled with no reason given.  This patch abstracts out the thread
canceling code and adds a reason for every cancellation.

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-threadupdate.c (cancel_thread): New.
(jump_thread_path_registry::thread_block_1): Use cancel_thread.
(jump_thread_path_registry::mark_threaded_blocks): Same.
(jump_thread_path_registry::register_jump_thread): Same.

3 years agoc++: Avoid bogus -Wunused with recent change
Jason Merrill [Thu, 2 Sep 2021 21:42:32 +0000 (17:42 -0400)]
c++: Avoid bogus -Wunused with recent change

My change to make limit_bad_template_recursion avoid instantiating members
of erroneous classes produced a bogus "used but not defined" warning for
23_containers/unordered_set/instantiation_neg.cc; it's not defined because
we decided not to instantiate it.  So we need to suppress that warning.

gcc/cp/ChangeLog:

* pt.c (limit_bad_template_recursion): Suppress -Wunused for decls
we decide not to instantiate.

3 years agoFortran: Fix Bind(C) char-len check, add ptr-contiguous check
Tobias Burnus [Fri, 3 Sep 2021 14:28:04 +0000 (16:28 +0200)]
Fortran: Fix Bind(C) char-len check, add ptr-contiguous check

Add F2018, 18.3.6 (5), pointer + contiguous is not permitted
check for dummies in BIND(C) procs.

Fix misreading of F2018, 18.3.4/18.3.5 + 18.3.6 (5) regarding
character dummies passed as byte stream to a bind(C) dummy arg:
Per F2018, 18.3.1 only len=1 is interoperable (since F2003).
F2008 added 'constant expression' for vars (F2018, 18.3.4/18.3.5),
applicable to dummy args per F2018, C1554. I misread this such
that len > 1 is permitted if len is a constant expr.

While the latter would work as character len=1 a(10) and len=2 a(5)
have the same storage sequence and len is fixed, it is still invalid.
Hence, it is now rejected again.

gcc/fortran/ChangeLog:

* decl.c (gfc_verify_c_interop_param): Reject pointer with
CONTIGUOUS attributes as dummy arg. Reject character len > 1
when passed as byte stream.

gcc/testsuite/ChangeLog:

* gfortran.dg/bind_c_char_6.f90: Update dg-error.
* gfortran.dg/bind_c_char_7.f90: Likewise.
* gfortran.dg/bind_c_char_8.f90: Likewise.
* gfortran.dg/iso_c_binding_char_1.f90: Likewise.
* gfortran.dg/pr32599.f03: Likewise.
* gfortran.dg/bind_c_char_9.f90: Comment testcase bits which are
implementable but not valid F2018.
* gfortran.dg/bind_c_contiguous.f90: New test.

3 years agoAvoid using unavailable objects in jt_state.
Aldy Hernandez [Fri, 3 Sep 2021 10:02:30 +0000 (12:02 +0200)]
Avoid using unavailable objects in jt_state.

The jump threading state is about to get more interesting, and it may
get with a ranger or with the const_copies/etc helpers.  This patch
makes sure we have an object before we attempt to call push_marker or
pop_to_marker.

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-threadedge.c (jt_state::push): Only call methods for
which objects are available.
(jt_state::pop): Same.
(jt_state::register_equiv): Same.
(jt_state::register_equivs_on_edge): Same.

3 years agoDo not release state location until after path registry.
Aldy Hernandez [Fri, 3 Sep 2021 08:11:28 +0000 (10:11 +0200)]
Do not release state location until after path registry.

We are popping state and then calling the registry code.  This causes
the registry to have incorrect information.  This isn't visible in
current trunk, but will be an issue when I submit further enhancements
to the threading code.  However, it is a cleanup on its own so I am
pushing it now.

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-threadedge.c (jump_threader::thread_across_edge):
Move pop until after a thread is registered.

3 years agoAdd debug helper for jump thread paths.
Aldy Hernandez [Fri, 3 Sep 2021 09:57:27 +0000 (11:57 +0200)]
Add debug helper for jump thread paths.

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-threadupdate.c (debug): New.

3 years agoRAII class to change dump_file.
Aldy Hernandez [Fri, 3 Sep 2021 09:45:03 +0000 (11:45 +0200)]
RAII class to change dump_file.

The function dump_ranger() shows everything the ranger knows at the
current time.  To do this, we tickle all the statements to force ranger
to provide as much information as possible.  During this process, the
relation code will dump status out to the dump_file, whereas in
dump_ranger, we want to dump it out to a specific file (most likely
stderr).  This patch changes the dump_file through the life of
dump_ranger() and resets it when its done.

This patch only affects dump/debugging code.

Tested on x86-64 Linux.

gcc/ChangeLog:

* gimple-range-trace.cc (push_dump_file::push_dump_file): New.
(push_dump_file::~push_dump_file): New.
(dump_ranger): Change dump_file temporarily while dumping
ranger.
* gimple-range-trace.h (class push_dump_file): New.

3 years agoAdd function name when dumping ranger contents.
Aldy Hernandez [Fri, 3 Sep 2021 09:32:11 +0000 (11:32 +0200)]
Add function name when dumping ranger contents.

These are minor cleanups to the dumping code.

Tested on x86-64 Linux.

gcc/ChangeLog:

* gimple-range-trace.cc (debug_seed_ranger): Remove static.
(dump_ranger): Dump function name.

3 years agoUse non-null knowledge in path_range_query.
Aldy Hernandez [Fri, 3 Sep 2021 09:22:26 +0000 (11:22 +0200)]
Use non-null knowledge in path_range_query.

This patch improves ranges for pointers we are interested in a path, by
using the non-null class from the ranger.  This allows us to thread more
paths with minimal effort.

Tested on x86-64 Linux.

gcc/ChangeLog:

* gimple-range-path.cc (path_range_query::range_defined_in_block):
Adjust for non-null.
(path_range_query::adjust_for_non_null_uses): New.
(path_range_query::precompute_ranges): Call
adjust_for_non_null_uses.
* gimple-range-path.h: Add m_non_null and
adjust_for_non_null_uses.

3 years agoImprove path_range_query dumps.
Aldy Hernandez [Fri, 3 Sep 2021 09:13:16 +0000 (11:13 +0200)]
Improve path_range_query dumps.

Tested on x86-64 Linux.

gcc/ChangeLog:

* gimple-range-path.cc (path_range_query::dump): Dump path
length.
(path_range_query::precompute_ranges): Dump entire path.

3 years agoImplement relation_oracle::debug.
Aldy Hernandez [Fri, 3 Sep 2021 08:58:27 +0000 (10:58 +0200)]
Implement relation_oracle::debug.

Tested on x86-64 Linux.

gcc/ChangeLog:

* value-relation.cc (relation_oracle::debug): New.
* value-relation.h (relation_oracle::debug): New.

3 years agoRemove unnecessary include from tree-ssa-loop-ch.c
Aldy Hernandez [Fri, 3 Sep 2021 08:47:27 +0000 (10:47 +0200)]
Remove unnecessary include from tree-ssa-loop-ch.c

Tested on x86-64 Linux.

gcc/ChangeLog:

* tree-ssa-loop-ch.c: Remove unnecessary include file.

3 years agoSkip statements with no BB in ranger.
Aldy Hernandez [Fri, 3 Sep 2021 08:42:37 +0000 (10:42 +0200)]
Skip statements with no BB in ranger.

The function postfold_gcond_edges() registers relations coming out of a
GIMPLE_COND.  With upcoming changes, we may be called with statements
not in the IL (for example, dummy statements created by the
forward threader).  This patch avoids breakage by exiting if the
statement does not have a defining basic block.  There is a similar
change to the path solver.

Tested on x86-64 Linux.

gcc/ChangeLog:

* gimple-range-fold.cc (fold_using_range::postfold_gcond_edges):
Skip statements with no defining BB.
* gimple-range-path.cc (path_range_query::range_defined_in_block):
Do not get confused by statements with no defining BB.

3 years agoImprove support for IMAGPART_EXPR and REALPART_EXPR in ranger.
Aldy Hernandez [Fri, 3 Sep 2021 07:58:29 +0000 (09:58 +0200)]
Improve support for IMAGPART_EXPR and REALPART_EXPR in ranger.

Currently we adjust statements containing an IMAGPART_EXPR if the
defining statement was one of a few built-ins known to return boolean
types.  We can also adjust statements for both IMAGPART_EXPR and
REALPART_EXPR where the defining statement is a constant.

This patch adds such support, and cleans up the code a bit.

Tested on x86-64 Linux.

gcc/ChangeLog:

* gimple-range-fold.cc (adjust_imagpart_expr): Move from
gimple_range_adjustment.  Add support for constants.
(adjust_realpart_expr): New.
(gimple_range_adjustment): Move IMAGPART_EXPR code to
adjust_imagpart_expr.
* range-op.cc (integral_table::integral_table): Add entry for
REALPART_CST.

3 years agolibgomp.*/error-1.{c,f90}: Fix dg-output newline pattern
Tobias Burnus [Fri, 3 Sep 2021 13:24:41 +0000 (15:24 +0200)]
libgomp.*/error-1.{c,f90}: Fix dg-output newline pattern

libgomp/ChangeLog:

* testsuite/libgomp.c-c++-common/error-1.c: Use \r\n not \n\r in
dg-output.
* testsuite/libgomp.fortran/error-1.f90: Likewise.

3 years agoImprove compatibility of -fdump-ada-spec with warnings
Eric Botcazou [Fri, 3 Sep 2021 09:18:01 +0000 (11:18 +0200)]
Improve compatibility of -fdump-ada-spec with warnings

This makes sure that the style and warning settings used in the
C/C++ bindings generated by -fdump-ada-spec do not leak into the
units that use them.

gcc/c-family/
* c-ada-spec.c (dump_ads): Generate pragmas to disable style checks
and -gnatwu warning for the package specification.

3 years agoopenmp: Improve expand_omp_atomic_pipeline
Jakub Jelinek [Fri, 3 Sep 2021 07:54:58 +0000 (09:54 +0200)]
openmp: Improve expand_omp_atomic_pipeline

When __atomic_* builtins were introduced, omp-expand.c (omp-low.c
at that point) has been adjusted in several spots so that it uses
the atomic builtins instead of sync builtins, but
expand_omp_atomic_pipeline has not because the __atomic_compare_exchange_*
APIs take address of the argument, so it kept using __sync_val_compare_swap_*.
That means it always uses seq_cst though.
This patch changes it to use the ATOMIC_COMPARE_EXCHANGE ifn which gimple-fold
folds __atomic_compare_exchange_* into - that ifn also passes expected
directly.

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

* omp-expand.c (expand_omp_atomic_pipeline): Use
IFN_ATOMIC_COMPARE_EXCHANGE instead of
BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_? so that memory order
can be provided.

3 years agoc++, abi: Set DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD on C++ zero width bitfields [PR102024]
Jakub Jelinek [Fri, 3 Sep 2021 07:46:32 +0000 (09:46 +0200)]
c++, abi: Set DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD on C++ zero width bitfields [PR102024]

The removal of remove_zero_width_bitfields function and its call from
C++ FE layout_class_type (which I've done in the P0466R5
layout-compatible helper intrinsics patch, so that the FE can actually
determine what is and isn't layout-compatible according to the spec)
unfortunately changed the ABI on various platforms.
The C FE has been keeping zero-width bitfields in the types, while
the C++ FE has been removing them after structure layout, so in various
cases when passing such structures in registers we had different ABI
between C and C++.

While both the C and C++ FE had some code to remove zero width bitfields
after structure layout, in both FEs it was buggy and didn't really remove
any.  In the C FE that code has been removed later on, while in the C++ FE
for GCC 4.5 in PR42217 it has been actually fixed, so the C++ FE started
to remove those bitfields.

The following patch doesn't change anything ABI-wise, but allows the
targets to decide what to do, emit -Wpsabi warnings etc.
Non-C zero width bitfields will be seen by the backends as normal
zero width bitfields, C++ zero width bitfields that used to be previously
removed will have DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD flag set.
I've reused the DECL_FIELD_ABI_IGNORED flag which is only used on non-bitfield
FIELD_DECLs right now, but the macros now check DECL_BIT_FIELD flag.

Each backend can then decide what it wants, whether it wants to keep
different ABI between C and C++ as in GCC 11 and older (i.e. incompatible
with G++ <= 4.4, compatible with G++ 4.5 .. 11), for that it would
ignore for the aggregate passing/returning decisions all
DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD FIELD_DECLs), whether it wants to never
ignore zero width bitfields (no changes needed for that case, except perhaps
-Wpsabi warning should be added and for that DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD
can be tested), or whether it wants to always ignore zero width bitfields
(I think e.g. riscv in GCC 10+ does that).

All this patch does is set the flag which the backends can then use.

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

PR target/102024
gcc/
* tree.h (DECL_FIELD_ABI_IGNORED): Changed into rvalue only macro
that is false if DECL_BIT_FIELD.
(SET_DECL_FIELD_ABI_IGNORED, DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD,
SET_DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD): Define.
* tree-streamer-out.c (pack_ts_decl_common_value_fields): For
DECL_BIT_FIELD stream DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD instead
of DECL_FIELD_ABI_IGNORED.
* tree-streamer-in.c (unpack_ts_decl_common_value_fields): Use
SET_DECL_FIELD_ABI_IGNORED instead of writing to
DECL_FIELD_ABI_IGNORED and for DECL_BIT_FIELD use
SET_DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD instead.
* lto-streamer-out.c (hash_tree): For DECL_BIT_FIELD hash
DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD instead of DECL_FIELD_ABI_IGNORED.
gcc/cp/
* class.c (build_base_field): Use SET_DECL_FIELD_ABI_IGNORED
instead of writing to DECL_FIELD_ABI_IGNORED.
(layout_class_type): Likewise.  In the place where zero-width
bitfields used to be removed, use
SET_DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD on those fields instead.
gcc/lto/
* lto-common.c (compare_tree_sccs_1): Also compare
DECL_FIELD_CXX_ZERO_WIDTH_BIT_FIELD values.

3 years agoRemove macro check for __AMX_BF16/INT8/TILE__ in header file.
liuhongt [Thu, 2 Sep 2021 04:49:46 +0000 (12:49 +0800)]
Remove macro check for __AMX_BF16/INT8/TILE__ in header file.

gcc/ChangeLog:

PR target/102166
* config/i386/amxbf16intrin.h : Remove macro check for __AMX_BF16__.
* config/i386/amxint8intrin.h : Remove macro check for __AMX_INT8__.
* config/i386/amxtileintrin.h : Remove macro check for __AMX_TILE__.

gcc/testsuite/ChangeLog:

PR target/102166
* g++.target/i386/pr102166.C: New test.

3 years agoDaily bump.
GCC Administrator [Fri, 3 Sep 2021 00:16:33 +0000 (00:16 +0000)]
Daily bump.

3 years agolibgfortran: Further fixes for GFC/CFI descriptor conversions.
Sandra Loosemore [Wed, 18 Aug 2021 14:22:03 +0000 (07:22 -0700)]
libgfortran: Further fixes for GFC/CFI descriptor conversions.

This patch is for:
PR100907 - Bind(c): failure handling wide character
PR100911 - Bind(c): failure handling C_PTR
PR100914 - Bind(c): errors handling complex
PR100915 - Bind(c): failure handling C_FUNPTR
PR100917 - Bind(c): errors handling long double real

All of these problems are related to the GFC descriptors constructed
by the Fortran front end containing ambigous or incomplete
information.  This patch does not attempt to change the GFC data
structure or the front end, and only makes the runtime interpret it in
more reasonable ways.  It's not a complete fix for any of the listed
issues.

The Fortran front end does not distinguish between C_PTR and
C_FUNPTR, mapping both onto BT_VOID.  That is what this patch does also.

The other bugs are related to GFC descriptors only containing elem_len
and not kind.  For complex types, the elem_len needs to be divided by
2 and then mapped onto a real kind.  On x86 targets, the kind
corresponding to C long double is different than its elem_len; since
we cannot accurately disambiguate between a 16-byte kind 10 long
double from __float128, this patch arbitrarily prefers to interpret that as
the standard long double type rather than the GNU extension.

Similarly, for character types, the GFC descriptor cannot distinguish
between character(kind=c_char, len=4) and character(kind=ucs4, len=1).
But since the front end currently rejects anything other than len=1
(PR92482) this patch uses the latter interpretation.

2021-09-01  Sandra Loosemore  <sandra@codesourcery.com>
    José Rui Faustino de Sousa  <jrfsousa@gmail.com>

gcc/testsuite/
PR fortran/100911
PR fortran/100915
PR fortran/100916
* gfortran.dg/PR100911.c: New file.
* gfortran.dg/PR100911.f90: New file.
* gfortran.dg/PR100914.c: New file.
* gfortran.dg/PR100914.f90: New file.
* gfortran.dg/PR100915.c: New file.
* gfortran.dg/PR100915.f90: New file.

libgfortran/
PR fortran/100907
PR fortran/100911
PR fortran/100914
PR fortran/100915
PR fortran/100917
* ISO_Fortran_binding-1-tmpl.h (CFI_type_cfunptr): Make equivalent
to CFI_type_cptr.
* runtime/ISO_Fortran_binding.c (cfi_desc_to_gfc_desc): Fix
handling of CFI_type_cptr and CFI_type_cfunptr.  Additional error
checking and code cleanup.
(gfc_desc_to_cfi_desc): Likewise.  Also correct kind mapping
for character, complex, and long double types.

3 years agoFortran: TS 29113 testsuite
Sandra Loosemore [Thu, 1 Jul 2021 03:03:27 +0000 (20:03 -0700)]
Fortran: TS 29113 testsuite

Add tests to exercise features added to Fortran via TS 29113, "Further
Interoperability of Fortran with C":

https://wg5-fortran.org/N1901-N1950/N1942.pdf

2021-09-01  Sandra Loosemore  <sandra@codesourcery.com>

gcc/testsuite/
* gfortran.dg/c-interop/allocatable-dummy-c.c: New file.
* gfortran.dg/c-interop/allocatable-dummy.f90: New file.
* gfortran.dg/c-interop/allocatable-optional-pointer.f90: New file.
* gfortran.dg/c-interop/allocate-c.c: New file.
* gfortran.dg/c-interop/allocate-errors-c.c: New file.
* gfortran.dg/c-interop/allocate-errors.f90: New file.
* gfortran.dg/c-interop/allocate.f90: New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-1.f90:
New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-2.f90:
New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-3.f90:
New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-4.f90:
New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-5.f90:
New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-6.f90:
New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-7.f90:
New file.
* gfortran.dg/c-interop/argument-association-assumed-rank-8.f90:
New file.
* gfortran.dg/c-interop/assumed-type-dummy.f90: New file.
* gfortran.dg/c-interop/c-interop.exp: New file.
* gfortran.dg/c-interop/c1255-1.f90: New file.
* gfortran.dg/c-interop/c1255-2.f90: New file.
* gfortran.dg/c-interop/c1255a.f90: New file.
* gfortran.dg/c-interop/c407a-1.f90: New file.
* gfortran.dg/c-interop/c407a-2.f90: New file.
* gfortran.dg/c-interop/c407b-1.f90: New file.
* gfortran.dg/c-interop/c407b-2.f90: New file.
* gfortran.dg/c-interop/c407c-1.f90: New file.
* gfortran.dg/c-interop/c516.f90: New file.
* gfortran.dg/c-interop/c524a.f90: New file.
* gfortran.dg/c-interop/c535a-1.f90: New file.
* gfortran.dg/c-interop/c535a-2.f90: New file.
* gfortran.dg/c-interop/c535b-1.f90: New file.
* gfortran.dg/c-interop/c535b-2.f90: New file.
* gfortran.dg/c-interop/c535b-3.f90: New file.
* gfortran.dg/c-interop/c535c-1.f90: New file.
* gfortran.dg/c-interop/c535c-2.f90: New file.
* gfortran.dg/c-interop/c535c-3.f90: New file.
* gfortran.dg/c-interop/c535c-4.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-1-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-1.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-2-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-2.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-3-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-3.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-4-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-4.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-5-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-5.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-6-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-6.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-7-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-7.f90: New file.
* gfortran.dg/c-interop/cf-descriptor-8-c.c: New file.
* gfortran.dg/c-interop/cf-descriptor-8.f90: New file.
* gfortran.dg/c-interop/cf-out-descriptor-1-c.c: New file.
* gfortran.dg/c-interop/cf-out-descriptor-1.f90: New file.
* gfortran.dg/c-interop/cf-out-descriptor-2-c.c: New file.
* gfortran.dg/c-interop/cf-out-descriptor-2.f90: New file.
* gfortran.dg/c-interop/cf-out-descriptor-3-c.c: New file.
* gfortran.dg/c-interop/cf-out-descriptor-3.f90: New file.
* gfortran.dg/c-interop/cf-out-descriptor-4-c.c: New file.
* gfortran.dg/c-interop/cf-out-descriptor-4.f90: New file.
* gfortran.dg/c-interop/cf-out-descriptor-5-c.c: New file.
* gfortran.dg/c-interop/cf-out-descriptor-5.f90: New file.
* gfortran.dg/c-interop/cf-out-descriptor-6-c.c: New file.
* gfortran.dg/c-interop/cf-out-descriptor-6.f90: New file.
* gfortran.dg/c-interop/contiguous-1-c.c: New file.
* gfortran.dg/c-interop/contiguous-1.f90: New file.
* gfortran.dg/c-interop/contiguous-2-c.c: New file.
* gfortran.dg/c-interop/contiguous-2.f90: New file.
* gfortran.dg/c-interop/contiguous-3-c.c: New file.
* gfortran.dg/c-interop/contiguous-3.f90: New file.
* gfortran.dg/c-interop/deferred-character-1.f90: New file.
* gfortran.dg/c-interop/deferred-character-2.f90: New file.
* gfortran.dg/c-interop/dump-descriptors.c: New file.
* gfortran.dg/c-interop/dump-descriptors.h: New file.
* gfortran.dg/c-interop/establish-c.c: New file.
* gfortran.dg/c-interop/establish-errors-c.c: New file.
* gfortran.dg/c-interop/establish-errors.f90: New file.
* gfortran.dg/c-interop/establish.f90: New file.
* gfortran.dg/c-interop/explicit-interface.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-1-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-1.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-2-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-2.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-3-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-3.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-4-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-4.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-5-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-5.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-6-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-6.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-7-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-7.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-8-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-8.f90: New file.
* gfortran.dg/c-interop/fc-descriptor-9-c.c: New file.
* gfortran.dg/c-interop/fc-descriptor-9.f90: New file.
* gfortran.dg/c-interop/fc-out-descriptor-1-c.c: New file.
* gfortran.dg/c-interop/fc-out-descriptor-1.f90: New file.
* gfortran.dg/c-interop/fc-out-descriptor-2-c.c: New file.
* gfortran.dg/c-interop/fc-out-descriptor-2.f90: New file.
* gfortran.dg/c-interop/fc-out-descriptor-3-c.c: New file.
* gfortran.dg/c-interop/fc-out-descriptor-3.f90: New file.
* gfortran.dg/c-interop/fc-out-descriptor-4-c.c: New file.
* gfortran.dg/c-interop/fc-out-descriptor-4.f90: New file.
* gfortran.dg/c-interop/fc-out-descriptor-5-c.c: New file.
* gfortran.dg/c-interop/fc-out-descriptor-5.f90: New file.
* gfortran.dg/c-interop/fc-out-descriptor-6-c.c: New file.
* gfortran.dg/c-interop/fc-out-descriptor-6.f90: New file.
* gfortran.dg/c-interop/fc-out-descriptor-7-c.c: New file.
* gfortran.dg/c-interop/fc-out-descriptor-7.f90: New file.
* gfortran.dg/c-interop/ff-descriptor-1.f90: New file.
* gfortran.dg/c-interop/ff-descriptor-2.f90: New file.
* gfortran.dg/c-interop/ff-descriptor-3.f90: New file.
* gfortran.dg/c-interop/ff-descriptor-4.f90: New file.
* gfortran.dg/c-interop/ff-descriptor-5.f90: New file.
* gfortran.dg/c-interop/ff-descriptor-6.f90: New file.
* gfortran.dg/c-interop/ff-descriptor-7.f90: New file.
* gfortran.dg/c-interop/note-5-3.f90: New file.
* gfortran.dg/c-interop/note-5-4-c.c: New file.
* gfortran.dg/c-interop/note-5-4.f90: New file.
* gfortran.dg/c-interop/optional-c.c: New file.
* gfortran.dg/c-interop/optional.f90: New file.
* gfortran.dg/c-interop/rank-class.f90: New file.
* gfortran.dg/c-interop/rank.f90: New file.
* gfortran.dg/c-interop/removed-restrictions-1.f90: New file.
* gfortran.dg/c-interop/removed-restrictions-2.f90: New file.
* gfortran.dg/c-interop/removed-restrictions-3.f90: New file.
* gfortran.dg/c-interop/removed-restrictions-4.f90: New file.
* gfortran.dg/c-interop/section-1-c.c: New file.
* gfortran.dg/c-interop/section-1.f90: New file.
* gfortran.dg/c-interop/section-1p.f90: New file.
* gfortran.dg/c-interop/section-2-c.c: New file.
* gfortran.dg/c-interop/section-2.f90: New file.
* gfortran.dg/c-interop/section-2p.f90: New file.
* gfortran.dg/c-interop/section-3-c.c: New file.
* gfortran.dg/c-interop/section-3.f90: New file.
* gfortran.dg/c-interop/section-3p.f90: New file.
* gfortran.dg/c-interop/section-4-c.c: New file.
* gfortran.dg/c-interop/section-4.f90: New file.
* gfortran.dg/c-interop/section-errors-c.c: New file.
* gfortran.dg/c-interop/section-errors.f90: New file.
* gfortran.dg/c-interop/select-c.c: New file.
* gfortran.dg/c-interop/select-errors-c.c: New file.
* gfortran.dg/c-interop/select-errors.f90: New file.
* gfortran.dg/c-interop/select.f90: New file.
* gfortran.dg/c-interop/setpointer-c.c: New file.
* gfortran.dg/c-interop/setpointer-errors-c.c: New file.
* gfortran.dg/c-interop/setpointer-errors.f90: New file.
* gfortran.dg/c-interop/setpointer.f90: New file.
* gfortran.dg/c-interop/shape.f90: New file.
* gfortran.dg/c-interop/size.f90: New file.
* gfortran.dg/c-interop/tkr.f90: New file.
* gfortran.dg/c-interop/typecodes-array-basic-c.c: New file.
* gfortran.dg/c-interop/typecodes-array-basic.f90: New file.
* gfortran.dg/c-interop/typecodes-array-char-c.c: New file.
* gfortran.dg/c-interop/typecodes-array-char.f90: New file.
* gfortran.dg/c-interop/typecodes-array-float128-c.c: New file.
* gfortran.dg/c-interop/typecodes-array-float128.f90: New file.
* gfortran.dg/c-interop/typecodes-array-int128-c.c: New file.
* gfortran.dg/c-interop/typecodes-array-int128.f90: New file.
* gfortran.dg/c-interop/typecodes-array-longdouble-c.c: New file.
* gfortran.dg/c-interop/typecodes-array-longdouble.f90: New file.
* gfortran.dg/c-interop/typecodes-sanity-c.c: New file.
* gfortran.dg/c-interop/typecodes-sanity.f90: New file.
* gfortran.dg/c-interop/typecodes-scalar-basic-c.c: New file.
* gfortran.dg/c-interop/typecodes-scalar-basic.f90: New file.
* gfortran.dg/c-interop/typecodes-scalar-float128-c.c: New file.
* gfortran.dg/c-interop/typecodes-scalar-float128.f90: New file.
* gfortran.dg/c-interop/typecodes-scalar-int128-c.c: New file.
* gfortran.dg/c-interop/typecodes-scalar-int128.f90: New file.
* gfortran.dg/c-interop/typecodes-scalar-longdouble-c.c: New file.
* gfortran.dg/c-interop/typecodes-scalar-longdouble.f90: New file.
* gfortran.dg/c-interop/ubound.f90: New file.
* lib/target-supports.exp
(check_effective_target_fortran_real_c_float128): New function.

3 years agolibstdc++: Implement std::atomic<T*>::compare_exchange_weak
Jonathan Wakely [Thu, 2 Sep 2021 15:47:31 +0000 (16:47 +0100)]
libstdc++: Implement std::atomic<T*>::compare_exchange_weak

For some reason r170217 didn't add compare_exchange_weak to the
__atomic_base<T*> partial specialization, and so weak compare exchange
operations on pointers use compare_exchange_strong instead.

This adds __atomic_base<T*>::compare_exchange_weak and then uses it in
std::atomic<T*>::compare_exchange_weak.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/atomic_base.h (__atomic_base<P*>::compare_exchange_weak):
Add new functions.
* include/std/atomic (atomic<T*>::compare_exchange_weak): Use
it.

3 years agolibstdc++: Tweak whitespace in <atomic>
Jonathan Wakely [Thu, 2 Sep 2021 15:08:25 +0000 (16:08 +0100)]
libstdc++: Tweak whitespace in <atomic>

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/std/atomic: Tweak whitespace.

3 years agolibstdc++: Remove "no stronger" assertion in compare exchange [PR102177]
Jonathan Wakely [Thu, 2 Sep 2021 14:29:22 +0000 (15:29 +0100)]
libstdc++: Remove "no stronger" assertion in compare exchange [PR102177]

P0418R2 removed some preconditions from std::atomic::compare_exchange_*
but we still enforce them via __glibcxx_assert. This removes those
assertions.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR c++/102177
* include/bits/atomic_base.h (__is_valid_cmpexch_failure_order):
New function to check if a memory order is valid for the failure
case of compare exchange operations.
(__atomic_base<I>::compare_exchange_weak): Simplify assertions
by using __is_valid_cmpexch_failure_order.
(__atomic_base<I>::compare_exchange_strong): Likewise.
(__atomic_base<P*>::compare_exchange_weak): Likewise.
(__atomic_base<P*>::compare_exchange_strong): Likewise.
(__atomic_impl::compare_exchange_weak): Add assertion.
(__atomic_impl::compare_exchange_strong): Likewise.
* include/std/atomic (atomic::compare_exchange_weak): Likewise.
(atomic::compare_exchange_strong): Likewise.

3 years agolibstdc++: Define std::invoke_r for C++23 (P2136R3)
Jonathan Wakely [Thu, 2 Sep 2021 10:54:12 +0000 (11:54 +0100)]
libstdc++: Define std::invoke_r for C++23 (P2136R3)

We already supported this feature as std::__invoke<R>, for internal use.
This just adds a public version of it to <functional>.

Internal uses should continue to include <bits/invoke.h> and use
std::__invoke<R> so that they don't need to include all of <functional>.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/std/functional (invoke_r): Define.
* include/std/version (__cpp_lib_invoke_r): Define.
* testsuite/20_util/function_objects/invoke/version.cc: Check
for __cpp_lib_invoke_r as well as __cpp_lib_invoke.
* testsuite/20_util/function_objects/invoke/4.cc: New test.

3 years agoImprove -Wuninitialized note location.
Martin Sebor [Thu, 2 Sep 2021 15:20:09 +0000 (09:20 -0600)]
Improve -Wuninitialized note location.

Related:
PR tree-optimization/17506 - warning about uninitialized variable points to wrong location
PR testsuite/37182 - Revision 139286 caused gcc.dg/pr17506.c and gcc.dg/uninit-15.c

gcc/ChangeLog:

PR tree-optimization/17506
PR testsuite/37182
* tree-ssa-uninit.c (warn_uninit): Remove conditional guarding note.

gcc/testsuite/ChangeLog:

PR tree-optimization/17506
PR testsuite/37182
* gcc.dg/diagnostic-tree-expr-ranges-2.c: Add expected output.
* gcc.dg/uninit-15-O0.c: Remove xfail.
* gcc.dg/uninit-15.c: Same.

3 years agoAdd support for device-modifiers for 'omp target device'.
Marcel Vollweiler [Thu, 2 Sep 2021 12:56:00 +0000 (05:56 -0700)]
Add support for device-modifiers for 'omp target device'.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/target-device-ancestor-4.f90: Comment out dg-final to avoid
UNRESOLVED.

3 years agoRefine fix for PR78185, improve LIM for code after inner loops
Richard Biener [Thu, 2 Sep 2021 08:47:35 +0000 (10:47 +0200)]
Refine fix for PR78185, improve LIM for code after inner loops

This refines the fix for PR78185 after understanding that the code
regarding to the comment 'In a loop that is always entered we may
proceed anyway.  But record that we entered it and stop once we leave
it.' was supposed to protect us from leaving possibly infinite inner
loops.  The simpler fix of moving the misplaced stopping code
can then be refined to continue processing when the exited inner
loop is finite, improving invariant motion for cases like in the
added testcase.

2021-09-02  Richard Biener  <rguenther@suse.de>

* tree-ssa-loop-im.c (fill_always_executed_in_1): Refine
fix for PR78185 and continue processing when leaving
finite inner loops.

* gcc.dg/tree-ssa/ssa-lim-16.c: New testcase.

3 years agomatch.pd: Demote IFN_{ADD,SUB,MUL}_OVERFLOW operands [PR99591]
Jakub Jelinek [Thu, 2 Sep 2021 09:25:07 +0000 (11:25 +0200)]
match.pd: Demote IFN_{ADD,SUB,MUL}_OVERFLOW operands [PR99591]

The overflow builtins work on infinite precision integers and then convert
to the result type's precision, so any argument promotions are useless.
The expand_arith_overflow expansion is able to demote the arguments itself
through get_range_pos_neg and get_min_precision calls and if needed promote
to whatever mode it decides to perform the operations in, but if there are
any promotions it demoted, those are already expanded.  Normally combine
would remove the useless sign or zero extensions when it sees the result
of those is only used in a lowpart subreg, but typically those lowpart
subregs appear multiple times in the pattern so that they describe properly
the overflow behavior and combine gives up, so we end up with e.g.
        movswl  %si, %esi
        movswl  %di, %edi
        imulw   %si, %di
        seto    %al
where both movswl insns are useless.

The following patch fixes it by demoting operands of the ifns (only gets
rid of integral to integral conversions that increase precision).
While IFN_{ADD,MUL}_OVERFLOW are commutative and just one simplify would be
enough, IFN_SUB_OVERFLOW is not, therefore two simplifications.

2021-09-02  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/99591
* match.pd: Demote operands of IFN_{ADD,SUB,MUL}_OVERFLOW if they
were promoted.

* gcc.target/i386/pr99591.c: New test.
* gcc.target/i386/pr97950.c: Match or reject setb or jn?b instructions
together with seta or jn?a.

3 years agoRevert "tree-optimization/102155 - fix LIM fill_always_executed_in CFG walk"
Richard Biener [Thu, 2 Sep 2021 07:58:39 +0000 (09:58 +0200)]
Revert "tree-optimization/102155 - fix LIM fill_always_executed_in CFG walk"

This reverts commit f482bf2af86990329b4df660f8c1eb9e094de9f9.

3 years agolibgcc: Enable hfmode soft-sf/df/xf/tf extensions and truncations.
liuhongt [Mon, 5 Jul 2021 09:31:46 +0000 (17:31 +0800)]
libgcc: Enable hfmode soft-sf/df/xf/tf extensions and truncations.

libgcc/ChangeLog:

* config/i386/32/sfp-machine.h (_FP_NANFRAC_H): New macro.
* config/i386/64/sfp-machine.h (_FP_NANFRAC_H): Ditto.
* config/i386/sfp-machine.h (_FP_NANSIGN_H): Ditto.
* config/i386/t-softfp: Add hf soft-fp.
* config.host: Add i386/64/t-softfp.
* config/i386/64/t-softfp: New file.

3 years agoEnable _Float16 type for TARGET_SSE2 and above.
liuhongt [Mon, 5 Jul 2021 09:05:45 +0000 (17:05 +0800)]
Enable _Float16 type for TARGET_SSE2 and above.

gcc/ChangeLog:

* config/i386/i386-modes.def (FLOAT_MODE): Define ieee HFmode.
* config/i386/i386.c (enum x86_64_reg_class): Add
X86_64_SSEHF_CLASS.
(merge_classes): Handle X86_64_SSEHF_CLASS.
(examine_argument): Ditto.
(construct_container): Ditto.
(classify_argument): Ditto, and set HFmode/HCmode to
X86_64_SSEHF_CLASS.
(function_value_32): Return _FLoat16/Complex Float16 by
%xmm0.
(function_value_64): Return _Float16/Complex Float16 by SSE
register.
(ix86_print_operand): Handle CONST_DOUBLE HFmode.
(ix86_secondary_reload): Require gpr as intermediate register
to store _Float16 from sse register when sse4 is not
available.
(ix86_libgcc_floating_mode_supported_p): Enable _FLoat16 under
sse2.
(ix86_scalar_mode_supported_p): Ditto.
(TARGET_LIBGCC_FLOATING_MODE_SUPPORTED_P): Defined.
* config/i386/i386.h (VALID_SSE2_REG_MODE): Add HFmode.
(VALID_INT_MODE_P): Add HFmode and HCmode.
* config/i386/i386.md (*pushhf_rex64): New define_insn.
(*pushhf): Ditto.
(*movhf_internal): Ditto.
* doc/extend.texi (Half-Precision Floating Point): Documemt
_Float16 for x86.

gcc/lto/ChangeLog:

* lto-lang.c (lto_type_for_mode): Return float16_type_node
when mode == TYPE_MODE (float16_type_node).

gcc/testsuite/ChangeLog

* gcc.target/i386/sse2-float16-1.c: New test.
* gcc.target/i386/sse2-float16-2.c: Ditto.
* gcc.target/i386/sse2-float16-3.c: Ditto.
* gcc.target/i386/float16-5.c: New test.

3 years agoUpdate hf soft-fp from glibc.
liuhongt [Wed, 7 Jul 2021 07:34:40 +0000 (15:34 +0800)]
Update hf soft-fp from glibc.

libgcc/ChangeLog

* soft-fp/eqhf2.c: New file.
* soft-fp/extendhfdf2.c: New file.
* soft-fp/extendhfsf2.c: New file.
* soft-fp/half.h (FP_CMP_EQ_H): New marco.
* soft-fp/truncdfhf2.c: New file
* soft-fp/truncsfhf2.c: New file

3 years agotree-optimization/102155 - fix LIM fill_always_executed_in CFG walk
Richard Biener [Wed, 1 Sep 2021 07:51:45 +0000 (09:51 +0200)]
tree-optimization/102155 - fix LIM fill_always_executed_in CFG walk

This fixes the CFG walk order of fill_always_executed_in to use
RPO oder rather than the dominator based order computed by
get_loop_body_in_dom_order.  That fixes correctness issues with
unordered dominator children.

The RPO order computed by rev_post_order_and_mark_dfs_back_seme in
its for-iteration mode is a good match for the algorithm.

2021-09-01  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102155
* tree-ssa-loop-im.c (fill_always_executed_in_1): Iterate
over a part of the RPO array and do not recurse here.
Dump blocks marked as always executed.
(fill_always_executed_in): Walk over the RPO array and
process loops whose header we run into.
(loop_invariant_motion_in_fun): Compute the first RPO
using rev_post_order_and_mark_dfs_back_seme in iteration
order and pass that to fill_always_executed_in.

3 years agoRevert "MIPS: add .module mipsREV to all output asm file"
YunQiang Su [Thu, 2 Sep 2021 02:17:18 +0000 (10:17 +0800)]
Revert "MIPS: add .module mipsREV to all output asm file"

This reverts commit d904008df267cbcc01bd6edf98fa0789fb6e94da.

".module MIPSRev" has higher priority than -march=octeon or like.
It makes assembler cannot recognize the extension instructions
of octeon (See pr62030-octeon.c).

3 years agoDaily bump.
GCC Administrator [Thu, 2 Sep 2021 00:16:59 +0000 (00:16 +0000)]
Daily bump.

3 years agoCall reduce_vector_comparison_to_scalar_comparison earlier
Jeff Law [Wed, 1 Sep 2021 23:13:58 +0000 (19:13 -0400)]
Call reduce_vector_comparison_to_scalar_comparison earlier

As noted in the PR, we can get an ICE after the introduction of code to reduce a vector comparison to a scalar.  The problem is we left the operand cache in an inconsistent state because we called the new function too late.   This is trivially fixed by making the transformation before we call update_stmt_if_modified.

The irony here is the whole point of calling reduce_vector_comparison_to_scalar_comparison when we did was to expose these kinds of secondary opportunities.  In this particular case we collapsed the test to a comparison of constants (thus no SSA operands).

Anyway, this fixes the problem in the obvious way.  This may all end up being moot if I can twiddle Richi's match.pd pattern to work.  It doesn't work as-written due to a couple issues that I haven't worked totally through yet.

Installed on the trunk after bootstrap & regression testing on x86 and verifying it addresses the aarch64 issue.

gcc/
PR tree-optimization/102152
* tree-ssa-dom.c (dom_opt_dom_walker::optimize_stmt): Reduce a vector
comparison to a scalar comparison before calling
update_stmt_if_modified.

gcc/testsuite/
PR tree-optimization/102152
* gcc.dg/pr102152.c: New test

3 years agoFix target/101934: aarch64 memset code creates unaligned stores for -mstrict-align
Andrew Pinski [Tue, 31 Aug 2021 04:41:14 +0000 (04:41 +0000)]
Fix target/101934: aarch64 memset code creates unaligned stores for -mstrict-align

The problem here is the aarch64_expand_setmem code did not check
STRICT_ALIGNMENT if it is creating an overlapping store.
This patch adds that check and the testcase works.

gcc/ChangeLog:

PR target/101934
* config/aarch64/aarch64.c (aarch64_expand_setmem):
Check STRICT_ALIGNMENT before creating an overlapping
store.

gcc/testsuite/ChangeLog:

PR target/101934
* gcc.target/aarch64/memset-strict-align-1.c: New test.

3 years agolibcpp: Implement C++23 P1949R7 - C++ Identifier Syntax using Unicode Standard Annex 31
Jakub Jelinek [Wed, 1 Sep 2021 20:33:06 +0000 (22:33 +0200)]
libcpp: Implement C++23 P1949R7 - C++ Identifier Syntax using Unicode Standard Annex 31

The following patch implements the
P1949R7 - C++ Identifier Syntax using Unicode Standard Annex 31
paper.  We already allow UTF-8 characters in the source, so that part
is already implemented, so IMHO all we need to do is pedwarn instead of
just warn for the (default) -Wnormalize=nfc (or for -Wnormalize={id,nkfc})
if the character is not in NFC and to use the unicode XID_Start and
XID_Continue derived code properties to find out what characters are allowed
(the standard actually adds U+005F to XID_Start, but we are handling the
ASCII compatible characters differently already and they aren't allowed
in UCNs in identifiers).  Instead of hardcoding the large tables
in ucnid.tab, this patch makes makeucnid.c read them from the Unicode
tables (13.0.0 version at this point).

For non-pedantic mode, we accept as 2nd+ char in identifiers a union
of valid characters in all supported modes, but for the 1st char it
was actually pedantically requiring that it is not any of the characters
that may not appear in the currently chosen standard as the first character.
This patch changes it such that also what is allowed at the start of an
identifier is a union of characters valid at the start of an identifier
in any of the pedantic modes.

2021-09-01  Jakub Jelinek  <jakub@redhat.com>

PR c++/100977
libcpp/
* include/cpplib.h (struct cpp_options): Add cxx23_identifiers.
* charset.c (CXX23, NXX23): New enumerators.
(CID, NFC, NKC, CTX): Renumber.
(ucn_valid_in_identifier): Implement P1949R7 - use CXX23 and
NXX23 flags for cxx23_identifiers.  For start character in
non-pedantic mode, allow characters that are allowed as start
characters in any of the supported language modes, rather than
disallowing characters allowed only as non-start characters in
current mode but for characters from other language modes allowing
them even if they are never allowed at start.
* init.c (struct lang_flags): Add cxx23_identifiers.
(lang_defaults): Add cxx23_identifiers column.
(cpp_set_lang): Initialize CPP_OPTION (pfile, cxx23_identifiers).
* lex.c (warn_about_normalization): If cxx23_identifiers, use
cpp_pedwarning_with_line instead of cpp_warning_with_line for
"is not in NFC" diagnostics.
* makeucnid.c: Adjust usage comment.
(CXX23, NXX23): New enumerators.
(all_languages): Add CXX23.
(not_NFC, not_NFKC, maybe_not_NFC): Renumber.
(read_derivedcore): New function.
(write_table): Print also CXX23 and NXX23 columns.
(main): Require 5 arguments instead of 4, call read_derivedcore.
* ucnid.h: Regenerated using Unicode 13.0.0 files.
gcc/testsuite/
* g++.dg/cpp23/normalize1.C: New test.
* g++.dg/cpp23/normalize2.C: New test.
* g++.dg/cpp23/normalize3.C: New test.
* g++.dg/cpp23/normalize4.C: New test.
* g++.dg/cpp23/normalize5.C: New test.
* g++.dg/cpp23/normalize6.C: New test.
* g++.dg/cpp23/normalize7.C: New test.
* g++.dg/cpp23/ucnid-1-utf8.C: New test.
* g++.dg/cpp23/ucnid-2-utf8.C: New test.
* gcc.dg/cpp/ucnid-4.c: Don't expect
"not valid at the start of an identifier" errors.
* gcc.dg/cpp/ucnid-4-utf8.c: Likewise.
* gcc.dg/cpp/ucnid-5-utf8.c: New test.

3 years agoAdd MIPS Linux support to gcc.misc-tests/linkage.c (testsuite/51748)
Andrew Pinski [Wed, 1 Sep 2021 00:13:51 +0000 (00:13 +0000)]
Add MIPS Linux support to gcc.misc-tests/linkage.c (testsuite/51748)

This adds MIPS Linux support to gcc.misc-tests/linkage.exp.  Basically
copying what was done for MIPS IRIX and changing the options to be correct.

OK?

gcc/testsuite/ChangeLog:

PR testsuite/51748
* gcc.misc-tests/linkage.exp: Add mips*-linux-* support.

3 years agoEnable ranger and caching in pass_waccess.
Martin Sebor [Wed, 1 Sep 2021 19:46:19 +0000 (13:46 -0600)]
Enable ranger and caching in pass_waccess.

gcc/ChangeLog:

* gimple-ssa-warn-access.cc (get_size_range): Add argument.
(check_access): Pass additional argument.
(check_memop_access): Remove template and make a member function.
(maybe_check_dealloc_call): Make a pass_waccess member function.
(class pass_waccess): Add, rename, and remove members.
(pass_waccess::pass_waccess): Adjust to name change.
(pass_waccess::~pass_waccess): Same.
(check_alloca): Make a member function.
(check_alloc_size_call): Same.
(check_strcat): Same.
(check_strncat): Same.
(check_stxcpy): Same.
(check_stxncpy): Same.
(check_strncmp): Same.
(maybe_warn_rdwr_sizes): Rename...
(pass_waccess::maybe_check_access_sizes): ...to this.
(pass_waccess::check_call): Adjust to name changes.
(pass_waccess::maybe_check_dealloc_call): Make a pass_waccess member
function.
(pass_waccess::execute): Adjust to name changes.
* gimple-ssa-warn-access.h (check_memop_access): Remove.
* pointer-query.cc (access_ref::phi): Handle null pointer.
(access_ref::inform_access): Same.
(pointer_query::put_ref): Modify a cached value, not a copy of it.
(pointer_query::dump): New function.
(compute_objsize_r): Avoid overwriting access_ref::bndrng.  Cache
more results.
* pointer-query.h (pointer_query::dump): Declare.
* tree-ssa-strlen.c (get_range): Simplify.  Use function query.
(dump_strlen_info): Use function query.
(printf_strlen_execute): Factor code out into pointer_query::put_ref.

gcc/testsuite/ChangeLog:

* gcc.dg/Wstringop-overflow-11.c: Remove xfails.
* gcc.dg/Wstringop-overflow-12.c: Same.
* gcc.dg/Wstringop-overflow-43.c: Add xfails.
* gcc.dg/Wstringop-overflow-73.c: New test.

3 years agolibcpp: __VA_OPT__ tweak
Jason Merrill [Wed, 1 Sep 2021 19:33:30 +0000 (21:33 +0200)]
libcpp: __VA_OPT__ tweak

> We want to remove the latter <placemarker> but not the former one, and
> the patch adds the vaopt_padding_tokens counter for it to control
> how many placemarkers are removed on vaopt_state::END.
> As can be seen in #c1 and #c2 of the PR, I've tried various approaches,
> but neither worked out for all the cases except the posted one.

I notice that the second placemarker you mention is avoid_paste, which seems
relevant.  This seems to also work, at least it doesn't seem to break any of
the va_opt tests.

2021-09-01  Jason Merrill  <jason@redhat.com>

* macro.c (replace_args): When __VA_OPT__ is on the LHS of ##,
remove trailing avoid_paste tokens.

3 years agolibcpp: __VA_OPT__ p1042r1 placemarker changes [PR101488]
Jakub Jelinek [Wed, 1 Sep 2021 19:31:25 +0000 (21:31 +0200)]
libcpp: __VA_OPT__ p1042r1 placemarker changes [PR101488]

So, besides missing #__VA_OPT__ patch for which I've posted patch last week,
P1042R1 introduced some placemarker changes for __VA_OPT__, most notably
the addition of before "removal of placemarker tokens," rescanning ...
and the
 #define H4(X, ...) __VA_OPT__(a X ## X) ## b
H4(, 1)  // replaced by a b
example mentioned there where we replace it currently with ab

The following patch are the minimum changes (except for the
__builtin_expect) that achieve the same preprocessing between current
clang++ and patched gcc on all the testcases I've tried (i.e. gcc __VA_OPT__
testsuite in c-c++-common/cpp/va-opt* including the new test and the clang
clang/test/Preprocessor/macro_va_opt* testcases).

At one point I was trying to implement the __VA_OPT__(args) case as if
for non-empty __VA_ARGS__ it expanded as if __VA_OPT__( and ) were missing,
but from the tests it seems that is not how it should work, in particular
if after (or before) we have some macro argument and it is not followed
(or preceded) by ##, then it should be macro expanded even when __VA_OPT__
is after ## or ) is followed by ##.  And it seems that not removing any
padding tokens isn't possible either, because the expansion of the arguments
typically has a padding token at the start and end and those at least
according to the testsuite need to go.  It is unclear if it would be enough
to remove just one or if all padding tokens should be removed.
Anyway, e.g. the previous removal of all padding tokens at the end of
__VA_OPT__ is undesirable, as it e.g. eats also the padding tokens needed
for the H4 example from the paper.

2021-09-01  Jakub Jelinek  <jakub@redhat.com>

PR preprocessor/101488
* macro.c (replace_args): Fix up handling of CPP_PADDING tokens at the
start or end of __VA_OPT__ arguments when preceeded or followed by ##.

* c-c++-common/cpp/va-opt-3.c: Adjust expected output.
* c-c++-common/cpp/va-opt-7.c: New test.

3 years agoc++: Add test for fixed PR [PR101592]
Marek Polacek [Wed, 1 Sep 2021 19:17:07 +0000 (15:17 -0400)]
c++: Add test for fixed PR [PR101592]

Fixed by my c++/99701 patch.

PR c++/101592

gcc/testsuite/ChangeLog:

* g++.dg/warn/Wlogical-op-3.C: New test.

3 years agolibiberty, configure, Darwin: Avoid detecting deprecated sbrk.
Iain Sandoe [Mon, 23 Aug 2021 16:27:25 +0000 (17:27 +0100)]
libiberty, configure, Darwin: Avoid detecting deprecated sbrk.

Darwin provides an implementation of sbrk, which is detected by
the configuration process.  However, it is deprecated which leads
to build warnings.  The malloc-based implementation is more
suitable.  This patch removes sbrk from the functions searched
for Darwin.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
libiberty/ChangeLog:

* configure: Regenerate.
* configure.ac: Do not search for sbrk on Darwin.
* xmalloc.c: Do not declare sbrk unless it has been found
by configure.

3 years agoFortran - improve wording of error message
Harald Anlauf [Wed, 1 Sep 2021 17:05:47 +0000 (19:05 +0200)]
Fortran - improve wording of error message

gcc/fortran/ChangeLog:

PR fortran/56985
* resolve.c (resolve_common_vars): Fix grammar and improve wording
of error message rejecting an unlimited polymorphic in COMMON.

3 years agoFix arm target build with inhibit_libc
Sebastian Huber [Mon, 30 Aug 2021 11:52:11 +0000 (13:52 +0200)]
Fix arm target build with inhibit_libc

Do not declare abort in "libgcc/unwind-arm-common.inc" since it is already
provided by "tsystem.h".  It fixes the following build error:

In file included from libgcc/config/arm/unwind-arm.c:144:
libgcc/unwind-arm-common.inc:55:24: error: macro "abort" passed 1 arguments, but takes just 0
   55 | extern void abort (void);

libgcc/

* unwind-arm-common.inc (abort): Remove.

3 years agoSimplify 'gcc/tree.c:walk_tree_1' handling of 'OMP_CLAUSE'
Thomas Schwinge [Fri, 27 Aug 2021 05:49:55 +0000 (07:49 +0200)]
Simplify 'gcc/tree.c:walk_tree_1' handling of 'OMP_CLAUSE'

No behavioral change, other than that for a few clauses, operands are now
walked in a different order, and 'OMP_CLAUSE_ERROR' now no longer runs into
'default: gcc_unreachable ();' here (but instead will at some later stage).

Follow-up for r110243 (commit aaf46ef9792bbc562175b606bd1c3f225ea56924)
"Fix PR 25886.  Convert OMP_CLAUSE_* into sub-codes".

gcc/
* tree.c (walk_tree_1) <OMP_CLAUSE>: Simplify.

3 years agoObjective-C, NeXT: Fix messenging non-aggregate return-in-memory.
Iain Sandoe [Mon, 16 Aug 2021 20:22:13 +0000 (21:22 +0100)]
Objective-C, NeXT: Fix messenging non-aggregate return-in-memory.

When a method returns a type that the platform ABI says should be
returned in memory, and that is done by a hidden 'sret' parameter,
the message send calls must be adjusted to inform the runtime that
the sret parameter is present.  As reported in the PR, this is not
working for non-aggregate types that use this mechanism.  The fix
here is to adjust the logic such that all return values that flag
'in memory' are considered to use the mechanism *unless* they
provide a struct_value_rtx *and* the return object is an aggregate.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
PR objc/101718 - Objective-C frontend emits wrong code to call methods returning scalar types returned in memory

PR objc/101718

gcc/objc/ChangeLog:

* objc-next-runtime-abi-02.c (build_v2_build_objc_method_call):
Revise for cases where scalar objects use an sret parameter.
(next_runtime_abi_02_build_objc_method_call): Likwise.

3 years agoC-family: Add attribute 'unavailable'.
Iain Sandoe [Sun, 8 Nov 2020 09:04:07 +0000 (09:04 +0000)]
C-family: Add attribute 'unavailable'.

If an interface is marked 'deprecated' then, presumably, at some point it
will be withdrawn and no longer available.  The 'unavailable' attribute
makes it possible to mark up interfaces to indicate this status.  It is used
quite extensively in some codebases where a single set of headers can be used
to permit code generation for multiple system versions.

From a configuration perspective, it also allows a compile test to determine
that an interface is missing - rather than requiring a link test.

The implementation follows the pattern of attribute deprecated, but produces
an error (where deprecation produces a warning).

This attribute has been implemented in clang for some years.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/c-family/ChangeLog:

* c-attribs.c (handle_unavailable_attribute): New.

gcc/c/ChangeLog:

* c-decl.c (enum deprecated_states): Add unavailable state.
(merge_decls): Copy unavailability.
(quals_from_declspecs): Handle unavailable case.
(start_decl): Amend the logic handling suppression of nested
deprecation states to include unavailability.
(smallest_type_quals_location): Amend comment.
(grokdeclarator): Handle the unavailable deprecation state.
(declspecs_add_type): Set TREE_UNAVAILABLE from the decl specs.
* c-tree.h (struct c_declspecs): Add unavailable_p.
* c-typeck.c (build_component_ref): Handle unavailability.
(build_external_ref): Likewise.

gcc/cp/ChangeLog:

* call.c (build_over_call): Handle unavailable state in addition to
deprecation.
* class.c (type_build_ctor_call): Likewise.
(type_build_dtor_call): Likewise.
* cp-tree.h: Rename cp_warn_deprecated_use to
cp_handle_deprecated_or_unavailable.
* decl.c (duplicate_decls): Merge unavailability.
(grokdeclarator): Handle unavailability in addition to deprecation.
(type_is_unavailable): New.
(grokparms): Handle unavailability in addition to deprecation.
* decl.h (enum deprecated_states): Add
UNAVAILABLE_DEPRECATED_SUPPRESS.
* decl2.c (cplus_decl_attributes): Propagate unavailability to
templates.
(cp_warn_deprecated_use): Rename to ...
(cp_handle_deprecated_or_unavailable): ... this and amend to handle
the unavailable case. It remains a warning in the case of deprecation
but becomes an error in the case of unavailability.
(cp_warn_deprecated_use_scopes): Handle unavailability.
(mark_used): Likewise.
* parser.c (cp_parser_template_name): Likewise.
(cp_parser_template_argument): Likewise.
(cp_parser_parameter_declaration_list): Likewise.
* typeck.c (build_class_member_access_expr): Likewise.
(finish_class_member_access_expr): Likewise.
* typeck2.c (build_functional_cast_1): Likewise.

gcc/ChangeLog:

* doc/extend.texi: Document unavailable attribute.
* print-tree.c (print_node): Handle unavailable attribute.
* tree-core.h (struct tree_base): Add a bit to carry unavailability.
* tree.c (error_unavailable_use): New.
* tree.h (TREE_UNAVAILABLE): New.
(error_unavailable_use): New.

gcc/objc/ChangeLog:

* objc-act.c (objc_add_property_declaration): Register unavailable
attribute.
(maybe_make_artificial_property_decl): Set available.
(objc_maybe_build_component_ref): Generalise to the method prototype
to count availability.
(objc_build_class_component_ref): Likewise.
(build_private_template): Likewise.
(objc_decl_method_attributes): Handle unavailable attribute.
(lookup_method_in_hash_lists): Amend comments.
(objc_finish_message_expr): Handle unavailability in addition to
deprecation.
(start_class): Likewise.
(finish_class): Likewise.
(lookup_protocol): Likewise.
(objc_declare_protocol): Likewise.
(start_protocol): Register unavailable attribute.
(really_start_method): Likewise.
(objc_gimplify_property_ref): Emit error on encountering an
unavailable entity (and a warning for a deprecated one).

gcc/testsuite/ChangeLog:

* g++.dg/ext/attr-unavailable-1.C: New test.
* g++.dg/ext/attr-unavailable-2.C: New test.
* g++.dg/ext/attr-unavailable-3.C: New test.
* g++.dg/ext/attr-unavailable-4.C: New test.
* g++.dg/ext/attr-unavailable-5.C: New test.
* g++.dg/ext/attr-unavailable-6.C: New test.
* g++.dg/ext/attr-unavailable-7.C: New test.
* g++.dg/ext/attr-unavailable-8.C: New test.
* g++.dg/ext/attr-unavailable-9.C: New test.
* gcc.dg/attr-unavailable-1.c: New test.
* gcc.dg/attr-unavailable-2.c: New test.
* gcc.dg/attr-unavailable-3.c: New test.
* gcc.dg/attr-unavailable-4.c: New test.
* gcc.dg/attr-unavailable-5.c: New test.
* gcc.dg/attr-unavailable-6.c: New test.
* obj-c++.dg/attributes/method-unavailable-1.mm: New test.
* obj-c++.dg/attributes/method-unavailable-2.mm: New test.
* obj-c++.dg/attributes/method-unavailable-3.mm: New test.
* obj-c++.dg/property/at-property-unavailable-1.mm: New test.
* obj-c++.dg/property/at-property-unavailable-2.mm: New test.
* obj-c++.dg/property/dotsyntax-unavailable-1.mm: New test.
* objc.dg/attributes/method-unavailable-1.m: New test.
* objc.dg/attributes/method-unavailable-2.m: New test.
* objc.dg/attributes/method-unavailable-3.m: New test.
* objc.dg/property/at-property-unavailable-1.m: New test.
* objc.dg/property/at-property-unavailable-2.m: New test.
* objc.dg/property/dotsyntax-unavailable-1.m: New test.

3 years agocoroutines : Add a missed begin/finish else clause to the codegen.
Iain Sandoe [Wed, 3 Mar 2021 16:14:24 +0000 (16:14 +0000)]
coroutines : Add a missed begin/finish else clause to the codegen.

Minor code-gen correction.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/cp/ChangeLog:

* coroutines.cc (build_actor_fn): Add begin/finish clauses
to the initial test in the actor function.

3 years agocoroutines: No cleanups on goto statements.
Iain Sandoe [Wed, 3 Mar 2021 16:13:00 +0000 (16:13 +0000)]
coroutines: No cleanups on goto statements.

Minor cleanup, this is statement not an expression, we do not
need to use finish_expr_stmt here.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/cp/ChangeLog:

* coroutines.cc (await_statement_walker): Use build_stmt and
add_stmt instead of build1 and finish_expr_stmt.

3 years agolibphobos: Update comment for DRUNTIME_OS_SOURCES
Iain Buclaw [Wed, 1 Sep 2021 12:18:16 +0000 (14:18 +0200)]
libphobos: Update comment for DRUNTIME_OS_SOURCES

libphobos/ChangeLog:

* m4/druntime/os.m4: Update comment for DRUNTIME_OS_SOURCES.

3 years agolibphobos: Don't add zlib when ENABLE_LIBDRUNTIME_ONLY
Iain Buclaw [Wed, 1 Sep 2021 12:11:05 +0000 (14:11 +0200)]
libphobos: Don't add zlib when ENABLE_LIBDRUNTIME_ONLY

The D run-time library does not depend on zlib, so only include it in
the library when Phobos is being built as well.

libphobos/ChangeLog:

* src/Makefile.am: Don't add zlib when ENABLE_LIBDRUNTIME_ONLY.
* src/Makefile.in: Regenerate.

3 years agovectorizer: Fix up vectorization using WIDEN_MINUS_EXPR [PR102124]
Jakub Jelinek [Wed, 1 Sep 2021 11:30:51 +0000 (13:30 +0200)]
vectorizer: Fix up vectorization using WIDEN_MINUS_EXPR [PR102124]

The following testcase is miscompiled on aarch64-linux at -O3 since the
introduction of WIDEN_MINUS_EXPR.
The problem is if the inner type (half_type) is unsigned and the result
type in which the subtraction is performed (type) has precision more than
twice as larger as the inner type's precision.
For other widening operations like WIDEN_{PLUS,MULT}_EXPR, if half_type
is unsigned, the addition/multiplication result in itype is also unsigned
and needs to be zero-extended to type.
But subtraction is special, even when half_type is unsigned, the subtraction
behaves as signed (also regardless of whether the result type is signed or
unsigned), 0xfeU - 0xffU is -1 or 0xffffffffU, not 0x0000ffff.

I think it is better not to use mixed signedness of types in
WIDEN_MINUS_EXPR (have unsigned vector of operands and signed result
vector), so this patch instead adds another cast to make sure we always
sign-extend the result from itype to type if type is wider than itype.

2021-09-01  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/102124
* tree-vect-patterns.c (vect_recog_widen_op_pattern): For ORIG_CODE
MINUS_EXPR, if itype is unsigned with smaller precision than type,
add an extra cast to signed variant of itype to ensure sign-extension.

* gcc.dg/torture/pr102124.c: New test.

3 years agograph output: use better colors for edges
Martin Liska [Wed, 1 Sep 2021 08:59:47 +0000 (10:59 +0200)]
graph output: use better colors for edges

gcc/ChangeLog:

* graph.c (draw_cfg_node_succ_edges): Do not color fallthru
  edges and rather use colors for TRUE and FALSE edges.

3 years agotree-optimization/93491 - avoid PRE of trapping calls across exits
Richard Biener [Wed, 1 Sep 2021 09:49:39 +0000 (11:49 +0200)]
tree-optimization/93491 - avoid PRE of trapping calls across exits

This makes us avoid PREing calls that could trap across other
calls that might not return.  The PR88087 testcase has exactly
such case so I've refactored the testcase to contain a valid PRE.
I've also adjusted PRE to not consider pure calls possibly
not returning in line with what we do elsewhere.

Note we don't have a good idea whether a function always returns
normally or whether its body is known to never trap.  That's
something IPA could compute.

2021-09-01  Richard Biener  <rguenther@suse.de>

PR tree-optimization/93491
* tree-ssa-pre.c (compute_avail): Set BB_MAY_NOTRETURN
after processing the stmt itself.  Do not consider
pure functions possibly not returning.  Properly avoid
adding possibly trapping calls to EXP_GEN when there's
a preceeding possibly not returning call.
* tree-ssa-sccvn.c (vn_reference_may_trap): Conservatively
not handle calls.

* gcc.dg/torture/pr93491.c: New testcase.
* gcc.dg/tree-ssa/pr88087.c: Change to valid PRE opportunity.

3 years agotree-optimization/102139 - fix SLP DR base alignment
Richard Biener [Tue, 31 Aug 2021 08:28:40 +0000 (10:28 +0200)]
tree-optimization/102139 - fix SLP DR base alignment

When doing whole-function SLP we have to make sure the recorded
base alignments we compute as the maximum alignment seen for a
base anywhere in the function is actually valid at the point
we want to make use of it.

To make this work we now record the stmt the alignment was derived
from in addition to the DRs innermost behavior and we use a
dominance check to verify the recorded info is valid when doing
BB vectorization.  For this to work for groups inside a BB that are
separate by a call that might not return we now store the DR
analysis group-id permanently and use that for an additional check
when the DRs are in the same BB.

2021-08-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102139
* tree-vectorizer.h (vec_base_alignments): Adjust hash-map
type to record a std::pair of the stmt-info and the innermost
loop behavior.
(dr_vec_info::group): New member.
* tree-vect-data-refs.c (vect_record_base_alignment): Adjust.
(vect_compute_data_ref_alignment): Verify the recorded
base alignment can be used.
(data_ref_pair): Remove.
(dr_group_sort_cmp): Adjust.
(vect_analyze_data_ref_accesses): Store the group-ID in the
dr_vec_info and operate on a vector of dr_vec_infos.

* gcc.dg/torture/pr102139.c: New testcase.

3 years agoMAINTAINERS: Add YunQiang Su for write after approval
YunQiang Su [Wed, 1 Sep 2021 10:19:12 +0000 (18:19 +0800)]
MAINTAINERS: Add YunQiang Su for write after approval

ChangeLog:
* MAINTAINERS: Add myself for write after approval.

3 years agomd/define_c_enum: support value assignation
YunQiang Su [Tue, 31 Aug 2021 11:19:49 +0000 (07:19 -0400)]
md/define_c_enum: support value assignation

Currently, the enums from define_c_enum and define_enum can only
has values one by one from 0.

In fact we can support the behaviour just like C, aka like
  (define_enum "mips_isa" [(mips1 1) mips2 (mips32 32) mips32r2]),
then we can get
  enum mips_isa {
    MIPS_ISA_MIPS1 = 1,
    MIPS_ISA_MIPS2 = 2,
    MIPS_ISA_MIPS32 = 32,
    MIPS_ISA_MIPS32R2 = 33
  };

gcc/ChangeLog:
* read-md.c (md_reader::handle_enum): support value assignation.
* doc/md.texi: record define_c_enum value assignation support.

3 years agobswap: Fix up bswap_view_convert handling [PR102141]
Jakub Jelinek [Wed, 1 Sep 2021 10:06:25 +0000 (12:06 +0200)]
bswap: Fix up bswap_view_convert handling [PR102141]

bswap_view_convert is used twice in spots where gsi_insert_before is the
right thing, but in the last one it wants to insert preparation stmts
for the VIEW_CONVERT_EXPR emitted with gsi_insert_after, where at the
gsi we still need to insert bswap_stmt and maybe mask_stmt whose lhs
the preparation stmts will use.
So, this patch adds a BEFORE argument to the function and emits the
preparation statements before or after depending on that.

2021-09-01  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/102141
* gimple-ssa-store-merging.c (bswap_view_convert): Add BEFORE
argument.  If false, emit stmts after gsi instead of before, and
with GSI_NEW_STMT.
(bswap_replace): Adjust callers.  When converting output of bswap,
emit VIEW_CONVERT prepratation stmts after a copy of gsi instead
of before it.

* gcc.dg/pr102141.c: New test.

3 years agotree-optimization/102149 - add testcase for fixed bug
Richard Biener [Wed, 1 Sep 2021 06:06:48 +0000 (08:06 +0200)]
tree-optimization/102149 - add testcase for fixed bug

This adds the testcase from the PR.

2021-09-01  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102149
* gcc.dg/torture/pr102149.c: New testcase.

3 years agoC: PR c/79412: Poison decls with error_mark_node after type mismatch
Roger Sayle [Wed, 1 Sep 2021 07:38:39 +0000 (08:38 +0100)]
C: PR c/79412: Poison decls with error_mark_node after type mismatch

This patch fixes an ICE during error-recovery regression in the C front-end.
The symptom is that the middle-end's sanity checking assertions fail during
gimplification when being asked to increment an array, which is non-sense.
The issue is that the C-front end has detected the type mismatch and
reported an error to the user, but hasn't provided any indication of this
to the middle-end, simply passing bogus trees that the optimizers recognize
as invalid.

This appears to be a frequently reported ICE with 94730, 94731, 101036
and 101365 all marked as duplicates.

I believe the correct (polite) fix is to mark the mismatched types as
problematic/dubious in the front-end, when the error is spotted, so that
the middle-end has a heads-up and can be a little more forgiving.  This
patch to c-decl.c's duplicate_decls sets (both) mismatched types to
error_mark_node if they are significantly different, and we've issued
an error message.  Alas, this is too punitive for FUNCTION_DECLs where
we store return types, parameter lists, parameter types and attributes
in the type, but fortunately the middle-end is already more cautious
about trusting possibly suspect function types.

This fix required one minor change to the testsuite, typedef-var-2.c
where after conflicting type definitions, we now no longer assume that
the (first or) second definition is the correct one.  This change only
affects the behaviour after seen_error(), so should be relatively safe.

2021-09-01  Roger Sayle  <roger@nextmovesoftware.com>
    Joseph Myers  <joseph@codesourcery.com>

gcc/c/ChangeLog
PR c/79412
* c-decl.c (duplicate_decls): On significant mismatches, mark the
types of both (non-function) decls as error_mark_node, so that the
middle-end can see the code is malformed.
(free_attr_access_data): Don't process if the type has been set to
error_mark_node.

gcc/testsuite/ChangeLog
PR c/79412
* gcc.dg/pr79412.c: New test case.
* gcc.dg/typedef-var-2.c: Update expeted errors.

3 years agoGet rid of all float-int special cases in validate_subreg.
liuhongt [Tue, 31 Aug 2021 09:36:52 +0000 (17:36 +0800)]
Get rid of all float-int special cases in validate_subreg.

gcc/ChangeLog:

* emit-rtl.c (validate_subreg): Get rid of all float-int
special cases.

3 years agoRevert "Make sure we're playing with integral modes before call extract_integral_bit_...
liuhongt [Tue, 31 Aug 2021 09:16:08 +0000 (17:16 +0800)]
Revert "Make sure we're playing with integral modes before call extract_integral_bit_field."

This reverts commit 7218c2ec365ce95f5a1012a6eb425b0a36aec6bf.

     PR middle-end/102133

3 years agoDaily bump.
GCC Administrator [Wed, 1 Sep 2021 00:16:58 +0000 (00:16 +0000)]
Daily bump.

3 years agoc++: Various small fixes
Jason Merrill [Wed, 25 Aug 2021 19:10:21 +0000 (15:10 -0400)]
c++: Various small fixes

A copy-paste error, a couple of missed checks to guard undefined accesses,
and we don't need to use type_uses_auto to extract the auto node we just
built.

gcc/cp/ChangeLog:

* coroutines.cc (flatten_await_stmt): Fix copyo.
* decl.c (reshape_init_class): Simplify.
* module.cc (module_state::read_language): Add null check.
* parser.c (build_range_temp): Avoid type_uses_auto.
(cp_parser_class_specifier_1): Add null check.

3 years agoFortran - extend set of substring expressions handled in length simplification
Harald Anlauf [Tue, 31 Aug 2021 19:00:53 +0000 (21:00 +0200)]
Fortran - extend set of substring expressions handled in length simplification

gcc/fortran/ChangeLog:

PR fortran/100950
* simplify.c (substring_has_constant_len): Minimize checks for
substring expressions being allowed.

gcc/testsuite/ChangeLog:

PR fortran/100950
* gfortran.dg/pr100950.f90: Extend coverage.

3 years agolibstdc++: Add valid range checks to std::span constructors [PR98421]
Jonathan Wakely [Tue, 31 Aug 2021 16:34:51 +0000 (17:34 +0100)]
libstdc++: Add valid range checks to std::span constructors [PR98421]

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR libstdc++/98421
* include/std/span (span(Iter, size_type), span(Iter, Iter)):
Add valid range checks.
* testsuite/23_containers/span/cons_1_assert_neg.cc: New test.
* testsuite/23_containers/span/cons_2_assert_neg.cc: New test.

3 years agoc++: check arity before deduction w/ explicit targs [PR12672]
Patrick Palka [Tue, 31 Aug 2021 17:31:10 +0000 (13:31 -0400)]
c++: check arity before deduction w/ explicit targs [PR12672]

During overload resolution, when the arity of a function template
clearly disagrees with the arity of the call, no specialization of the
function template could yield a viable candidate.  The deduction routine
type_unification_real already notices this situation, but not before
it substitutes explicit template arguments into the template, a step
which could induce a hard error.  Although it's necessary to perform
this substitution first in order to check arity perfectly (since the
substitution can e.g. expand a non-trailing parameter pack), in most
cases we can determine ahead of time whether there's an arity
disagreement without needing to perform deduction at all.

To that end, this patch implements an (approximate) arity check in
add_template_candidate_real that guards actual deduction.  It's enabled
only when there are explicit template arguments since that's when
deduction can force otherwise avoidable template instantiations.  (I
experimented with enabling it unconditionally as an optimization, and
observed some improvements to compile time of about 5% but also some
slowdowns of about the same magnitude, so kept it conditional.)

In passing, this adds a least_p parameter to arity_rejection for sake
of consistent diagnostics with unify_arity.

A couple of testcases needed to be adjusted so that deduction continues
to occur as intended after this change.  Except in unify6.C, where we
were expecting foo<void ()> to be ill-formed due to substitution
forming a function type with an added 'const', but ISTM this is
permitted by [dcl.fct]/7, so I changed the test accordingly.

PR c++/12672

gcc/cp/ChangeLog:

* call.c (rejection_reason::call_varargs_p): Rename this
previously unused member to ...
(rejection_reason::least_p): ... this.
(arity_rejection): Add least_p parameter.
(add_template_candidate_real): When there are explicit
template arguments, check that the arity of the call agrees with
the arity of the function before attempting deduction.
(print_arity_information): Add least_p parameter.
(print_z_candidate): Adjust call to print_arity_information.

gcc/testsuite/ChangeLog:

* g++.dg/cpp0x/decltype29.C: Adjust.
* g++.dg/template/error56.C: Adjust.
* g++.old-deja/g++.pt/unify6.C: Adjust.
* g++.dg/template/explicit-args7.C: New test.

3 years agoFix 'OMP_CLAUSE_TILE' operands handling in 'gcc/tree.c:walk_tree_1'
Thomas Schwinge [Fri, 27 Aug 2021 05:49:35 +0000 (07:49 +0200)]
Fix 'OMP_CLAUSE_TILE' operands handling in 'gcc/tree.c:walk_tree_1'

In r245300 (commit 02889d23ee3b02854dff203dd87b9a25e30b61b4)
"OpenACC tile clause support" that one had changed to three operands,
similar to 'OMP_CLAUSE_COLLAPSE'.

There is no (existing) test case where this seems to matter (likewise
for 'OMP_CLAUSE_COLLAPSE'), but it's good to be consistent.

gcc/
* tree.c (walk_tree_1) <OMP_CLAUSE_TILE>: Handle three operands.

3 years ago[OMP] Standardize on 'omp_privatize_by_reference'
Thomas Schwinge [Wed, 29 May 2019 16:59:46 +0000 (18:59 +0200)]
[OMP] Standardize on 'omp_privatize_by_reference'

... instead of 'omp_is_reference' vs.
'lang_hooks.decls.omp_privatize_by_reference'.

gcc/
* omp-general.h (omp_is_reference): Rename to...
(omp_privatize_by_reference): ... this.  Adjust all users...
* omp-general.c: ... here, ...
* gimplify.c: ... here, ...
* omp-expand.c: ... here, ...
* omp-low.c: ... here.

3 years agoAvoid valid Coverity warning for comparing array to zero.
Martin Sebor [Tue, 31 Aug 2021 17:16:37 +0000 (11:16 -0600)]
Avoid valid Coverity warning for comparing array to zero.

* gimple-ssa-warn-access.cc (maybe_warn_alloc_args_overflow): Test
pointer element for equality to zero, not that of the cotaining
array.

3 years agoDisable gcc_rich_location copying and assignment.
Martin Sebor [Tue, 31 Aug 2021 17:15:21 +0000 (11:15 -0600)]
Disable gcc_rich_location copying and assignment.

gcc/cp/ChangeLog:

* parser.c (cp_parser_selection_statement): Use direct initialization
instead of copy.

gcc/ChangeLog:

* gcc-rich-location.h (gcc_rich_location): Make ctor explicit.

libcpp/ChangeLog:

* include/line-map.h (class rich_location): Disable copying and
assignment.

3 years agoAdd attribute returns nonnull to get_range_query.
Martin Sebor [Tue, 31 Aug 2021 17:13:29 +0000 (11:13 -0600)]
Add attribute returns nonnull to get_range_query.

gcc/ChangeLog:
* function.h (function): Add comments.
(get_range_query): Same.  Add attribute returns nonnull.

3 years agolibstdc++: Fix broken autoconf check for O_NONBLOCK
Jonathan Wakely [Tue, 31 Aug 2021 16:08:00 +0000 (17:08 +0100)]
libstdc++: Fix broken autoconf check for O_NONBLOCK

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* configure.ac: Fix checks for F_GETFL, F_SETFL and O_NONBLOCK.
* configure: Regenerate.

3 years agolibstdc++: Remove redundant noexcept-specifier on definitions
Jonathan Wakely [Tue, 31 Aug 2021 15:30:01 +0000 (16:30 +0100)]
libstdc++: Remove redundant noexcept-specifier on definitions

These destructors are noexcept anyway. I removed the redundant noexcept
from the error_category destructor's declaration in r0-123475, but
didn't remove it from the defaulted definition in system_error.cc. That
causes warnings if the library is built with Clang.

This removes the redundant noexcept from ~error_category and
~system_error and adds tests to ensure they really are noexcept.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* src/c++11/system_error.cc (error_category::~error_category()):
Remove noexcept-specifier.
(system_error::~system_error()): Likewise.
* testsuite/19_diagnostics/error_category/noexcept.cc: New test.
* testsuite/19_diagnostics/system_error/noexcept.cc: New test.

3 years agolibstdc++: Add missing return for atomic timed wait [PR102074]
Jonathan Wakely [Tue, 31 Aug 2021 15:50:17 +0000 (16:50 +0100)]
libstdc++: Add missing return for atomic timed wait [PR102074]

This adds a missing return statement to the non-futex wait-until
operation.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR libstdc++/102074
* include/bits/atomic_timed_wait.h (__timed_waiter_pool)
[!_GLIBCXX_HAVE_PLATFORM_TIMED_WAIT]: Add missing return.

3 years agolibstdc++: Improve error handling in Net TS name resolution
Jonathan Wakely [Tue, 31 Aug 2021 12:09:26 +0000 (13:09 +0100)]
libstdc++: Improve error handling in Net TS name resolution

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/experimental/internet (__make_resolver_error_code):
Handle EAI_SYSTEM errors.
(basic_resolver_results): Use __make_resolver_error_code. Use
Glibc NI_MAXHOST and NI_MAXSERV values for buffer sizes.

3 years agolibstdc++: Fix ip::tcp::resolver test failure on Solaris
Jonathan Wakely [Tue, 31 Aug 2021 12:08:23 +0000 (13:08 +0100)]
libstdc++: Fix ip::tcp::resolver test failure on Solaris

Solaris 11 does not have "http" in /etc/services, which causes this test
to fail. Try some other services until we find one that works.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* testsuite/experimental/net/internet/resolver/ops/lookup.cc:
Try other service if "http" fails.

3 years ago[Committed] Fix subreg_promoted_mode breakage on various platforms.
Roger Sayle [Tue, 31 Aug 2021 16:25:58 +0000 (17:25 +0100)]
[Committed] Fix subreg_promoted_mode breakage on various platforms.

My apologies for the inconvenience.  My recent patch to preserve
SUBREG_PROMOTED_VAR_P on (extend:HI (subreg/s:QI (reg:SI))), and other
places in the middle-end, has broken the build on several targets.

The change to convert_modes inadvertently used the same
subreg_promoted_mode idiom for retrieving the mode of a SUBREG_REG
as the existing code just a few lines earlier.  Alas in the meantime,
the original SUBREG gets replaced by one without SUBREG_PROMOTED_VAR_P,
the whole raison-d'etre for my patch, and I'd not realized/noticed
that subreg_promoted_mode asserts for this.  Alas neither the bootstrap
and regression test on x86_64-pc-linux-gnu nor my testing on nvptx-none
must have hit this particular case.  The logic of this transformation
is sound, it's the implementation that's bitten me.

This patch has been committed, after another "make bootstrap" on
x86_64-pc-linux-gnu (just in case), and confirmation/pre-approval
from Jeff Law that this indeed fixes the build failures seen on
several platforms.

My humble apologies again.

2021-08-31  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* expr.c (convert_modes): Don't use subreg_promoted_mode on a
SUBREG if it can't be guaranteed to a SUBREG_PROMOTED_VAR_P set.
Instead use the standard (safer) is_a <scalar_int_mode> idiom.

3 years agoc++: use iloc_sentinel in another place
Jason Merrill [Sat, 28 Aug 2021 02:59:48 +0000 (22:59 -0400)]
c++: use iloc_sentinel in another place

Another place we can use iloc_sentinel instead of explicitly saving and
restoring input_location.

gcc/cp/ChangeLog:

* constexpr.c (explain_invalid_constexpr_fn): Use iloc_sentinel.

3 years agoc++: Improve error recovery with constexpr [PR92193]
Jason Merrill [Mon, 30 Aug 2021 22:42:05 +0000 (18:42 -0400)]
c++: Improve error recovery with constexpr [PR92193]

The compiler tries to limit error cascades in limit_bad_template_recursion
by avoiding triggering a new instantiation from one that has caused errors.
We were exempting constexpr functions from this because they can be needed
for constant evaluation, but as more and more functions get marked
constexpr, this becomes an over-broad category.  So as suggested on IRC,
this patch only exempts functions that are needed for mandatory constant
evaluation.

As noted in the comment, this flag doesn't particularly need to use a bit in
the FUNCTION_DECL, but there were still some free.

PR c++/92193

gcc/cp/ChangeLog:

* cp-tree.h (FNDECL_MANIFESTLY_CONST_EVALUATED): New.
* constexpr.c (cxx_eval_call_expression): Set it.
* pt.c (neglectable_inst_p): Check it.

gcc/testsuite/ChangeLog:

* g++.dg/diagnostic/static_assert4.C: New test.

3 years agoRestore intent of data-sym-multi-pool test
Jeff Law [Tue, 31 Aug 2021 15:08:50 +0000 (11:08 -0400)]
Restore intent of data-sym-multi-pool test

gcc/testsuite
* gcc.target/mips/mips.exp: Add tree-vrp to mips_option_group.
* gcc.target/mips/data-sym-multi-pool.c: Add -fno-tree-vrp.

3 years agoMore stabs removal.
Jeff Law [Tue, 31 Aug 2021 13:48:02 +0000 (09:48 -0400)]
More stabs removal.

gcc/

* config.gcc (cris-*-elf, cris-*-none): Remove dbxelf.h from
tm_file.
(m32r-*-elf, m32rle-*-elf, m32r-*-linux): Likewise.
(mn10300-*-*, am33_2.0-*-linux*): Likewise.
(xtensa*-*-elf, xtensa*-*-linux, xtensa*-*-uclinux): Likewise.
(m32c-*-elf*, m32c-*-rtems*): Likewise.
* config/cris/cris.h (DBX_NO_XREFS): Remove.
(DBX_CONTIN_LENGTH, DBX_CONTIN_CHAR): Likewise.
* config/m32r/m32r.h (DBXOUT_SOURCE_LINE): Likewise.
(DBX_DEBUGGING_INFO, DBX_CONTIN_LENGTH): Likewise.
* config/mn10300/mn10300.h (DEFAULT_GDB_EXTENSIONS): Likewise.
* config/mn10300/linux.h (DBX_REGISTER_NAMES): Likewise.

3 years agotestsuite: Fix gcc.dg/vect/pr101145* tests [PR101145]
Jakub Jelinek [Tue, 31 Aug 2021 13:26:14 +0000 (15:26 +0200)]
testsuite: Fix gcc.dg/vect/pr101145* tests [PR101145]

I'm getting:
FAIL: gcc.dg/vect/pr101145.c scan-tree-dump-times vect "vectorized 1 loops" 7
FAIL: gcc.dg/vect/pr101145_1.c scan-tree-dump-times vect "vectorized 1 loops" 2
FAIL: gcc.dg/vect/pr101145_2.c scan-tree-dump-times vect "vectorized 1 loops" 2
FAIL: gcc.dg/vect/pr101145_3.c scan-tree-dump-times vect "vectorized 1 loops" 2
FAIL: gcc.dg/vect/pr101145.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 1 loops" 7
FAIL: gcc.dg/vect/pr101145_1.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 1 loops" 2
FAIL: gcc.dg/vect/pr101145_2.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 1 loops" 2
FAIL: gcc.dg/vect/pr101145_3.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 1 loops" 2
on i686-linux (or x86_64-linux with -m32/-mno-sse).
The problem is that those tests use dg-options, which in */vect/ testsuite
throws away all the carefully added default options to enable vectorization
on each target (and which e.g. vect_int etc. effective targets rely on).
The old way would be to name those tests gcc.dg/vect/O3-pr101145*,
but we can also use dg-additional-options (which doesn't throw the default
options, just appends to them) which is IMO better so that we don't have to
rename the tests.

2021-08-31  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/101145
* gcc.dg/vect/pr101145.c: Use dg-additional-options with just -O3
instead of dg-options with -O3 -fdump-tree-vect-details.
* gcc.dg/vect/pr101145_1.c: Likewise.
* gcc.dg/vect/pr101145_2.c: Likewise.
* gcc.dg/vect/pr101145_3.c: Likewise.

3 years agoAdd support for device-modifiers for 'omp target device'.
Marcel Vollweiler [Tue, 31 Aug 2021 13:09:40 +0000 (06:09 -0700)]
Add support for device-modifiers for 'omp target device'.

'device_num' and 'ancestor' are now parsed on target device constructs for C,
C++, and Fortran (see OpenMP specification 5.0, p. 170). When 'ancestor' is
 used, then 'sorry, not supported' is output. Moreover, the restrictions for
'ancestor' are implemented (see OpenMP specification 5.0, p. 174f).

gcc/c/ChangeLog:

* c-parser.c (c_parser_omp_clause_device): Parse device-modifiers 'device_num'
and 'ancestor' in 'target device' clauses.

gcc/cp/ChangeLog:

* parser.c (cp_parser_omp_clause_device): Parse device-modifiers 'device_num'
and 'ancestor' in 'target device' clauses.
* semantics.c (finish_omp_clauses): Error handling. Constant device ids must
evaluate to '1' if 'ancestor' is used.

gcc/fortran/ChangeLog:

* gfortran.h: Add variable for 'ancestor' in struct gfc_omp_clauses.
* openmp.c (gfc_match_omp_clauses): Parse device-modifiers 'device_num'
and 'ancestor' in 'target device' clauses.
* trans-openmp.c (gfc_trans_omp_clauses): Set OMP_CLAUSE_DEVICE_ANCESTOR.

gcc/ChangeLog:

* gimplify.c (gimplify_scan_omp_clauses): Error handling. 'ancestor' only
allowed on target constructs and only with particular other clauses.
* omp-expand.c (expand_omp_target): Output of 'sorry, not supported' if
'ancestor' is used.
* omp-low.c (check_omp_nesting_restrictions): Error handling. No nested OpenMP
structs when 'ancestor' is used.
(scan_omp_1_stmt): No usage of OpenMP runtime routines in a target region when
'ancestor' is used.
* tree-pretty-print.c (dump_omp_clause): Append 'ancestor'.
* tree.h (OMP_CLAUSE_DEVICE_ANCESTOR): Define macro.

gcc/testsuite/ChangeLog:

* c-c++-common/gomp/target-device-1.c: New test.
* c-c++-common/gomp/target-device-2.c: New test.
* c-c++-common/gomp/target-device-ancestor-1.c: New test.
* c-c++-common/gomp/target-device-ancestor-2.c: New test.
* c-c++-common/gomp/target-device-ancestor-3.c: New test.
* c-c++-common/gomp/target-device-ancestor-4.c: New test.
* gfortran.dg/gomp/target-device-1.f90: New test.
* gfortran.dg/gomp/target-device-2.f90: New test.
* gfortran.dg/gomp/target-device-ancestor-1.f90: New test.
* gfortran.dg/gomp/target-device-ancestor-2.f90: New test.
* gfortran.dg/gomp/target-device-ancestor-3.f90: New test.
* gfortran.dg/gomp/target-device-ancestor-4.f90: New test.

3 years agolibstdc++: Fix 17_intro/names.cc failures on Solaris
Jonathan Wakely [Tue, 31 Aug 2021 09:25:53 +0000 (10:25 +0100)]
libstdc++: Fix 17_intro/names.cc failures on Solaris

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* testsuite/17_intro/names.cc: Undefine some more names used
by Solaris system headers.

3 years agoPreserve SUBREG_PROMOTED_VAR_P on (extend:HI (subreg/s:QI (reg:SI))).
Roger Sayle [Tue, 31 Aug 2021 10:45:07 +0000 (11:45 +0100)]
Preserve SUBREG_PROMOTED_VAR_P on (extend:HI (subreg/s:QI (reg:SI))).

SUBREG_PROMOTED_VAR_P is a mechanism for tracking that a partial subreg
is correctly zero-extended or sign-extended in the parent register.  For
example, the RTL (subreg/s/v:QI (reg/v:SI 23 [ x ]) 0) indicates that the
byte x is zero extended in reg:SI 23, which is useful for optimization.
An example is that zero extending the above QImode value to HImode can
simply use a wider subreg, i.e. (subreg:HI (reg/v:SI 23 [ x ]) 0).

This patch addresses the oversight/missed optimization opportunity that
the new HImode subreg above should retain its SUBREG_PROMOTED_VAR_P
annotation as its value is guaranteed to be correctly extended in the
SImode parent.  The code below to preserve SUBREG_PROMOTED_VAR_P is already
present in the middle-end (e.g. simplify-rtx.c:7232-7242) but missing
from one or two (precisely three) places that (accidentally) strip it.

Whilst there I also added another optimization.  If we need to extend
the above QImode value beyond the SImode register holding it, say to
DImode, we can eliminate the SUBREG and simply extend from the SImode
register to DImode.

2021-08-31  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* expr.c (convert_modes): Preserve SUBREG_PROMOTED_VAR_P when
creating a (wider) partial subreg from a SUBREG_PROMOTED_VAR_P
subreg.
* simplify-rtx.c (simplify_unary_operation_1) [SIGN_EXTEND]:
Likewise, preserve SUBREG_PROMOTED_VAR_P when creating a (wider)
partial subreg from a SUBREG_PROMOTED_VAR_P subreg.  Generate
SIGN_EXTEND of the SUBREG_REG when a subreg would be paradoxical.
[ZERO_EXTEND]: Likewise, preserve SUBREG_PROMOTED_VAR_P when
creating a (wider) partial subreg from a SUBREG_PROMOTED_VAR_P
subreg.  Generate ZERO_EXTEND of the SUBREG_REG when a subreg
would be paradoxical.

3 years agoOnly simplify TRUNCATE to SUBREG on TRULY_NOOP_TRUNCATION targets.
Roger Sayle [Tue, 31 Aug 2021 10:41:57 +0000 (11:41 +0100)]
Only simplify TRUNCATE to SUBREG on TRULY_NOOP_TRUNCATION targets.

As recently remarked by Jeff Law, SUBREGs are the "forever chemicals"
of GCC's RTL; once created they persist in the environment.  The problem,
according to the comment on lines 5428-5438 of combine.c is that
non-tieable SUBREGs interfere with reload/register allocation, so
combine often doesn't touch/clean-up instructions containing a SUBREG.

This is the first and simplest of two patches to tackle that problem,
by teaching combine to avoid converting explicit TRUNCATEs into
SUBREGs that it can't handle.

Consider the following (hypothetical) sequence of instructions on
a STORE_FLAG_VALUE=1 target, which stores a zero or one in an SI
register, then uselessly truncates to QImode, then extends it again.

(set (reg:SI 27) (ne:SI (reg:BI 28) (const_int 0)))
(set (reg:QI 26) (truncate:QI (reg:SI 27)))
(set (reg:SI 0) (zero_extend:SI (reg:QI 26)))

which ideally (i.e. with this patch) combine would simplify to:
(set (reg:SI 0) (ne:SI (reg:BI 28) (const_int 0)))

Alas currently, during combine the middle TRUNCATE is converted into
a lowpart SUBREG, which subst then turns into (clobber (const_int 0)),
abandoning the attempted combination, that then never reaches recog.

2021-08-31  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* combine.c (combine_simplify_rtx): Avoid converting an explicit
TRUNCATE into a lowpart SUBREG on !TRULY_NOOP_TRUNCATION targets.
* simplify-rtx.c (simplify_unary_operation_1): Likewise.

3 years agotree-optimization/102142 - fix typo in loop BB reduc cost adjustment
Richard Biener [Tue, 31 Aug 2021 09:04:51 +0000 (11:04 +0200)]
tree-optimization/102142 - fix typo in loop BB reduc cost adjustment

This fixes a typo in the condition guarding the cleanup of the
visited flag of costed scalar stmts.

2021-08-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102142
* tree-vect-slp.c (vect_bb_vectorization_profitable_p): Fix
condition under which to unset the visited flag.

* g++.dg/torture/pr102142.C: New testcase.

3 years agolibgcc: Add missing runtime exception notices
Richard Sandiford [Tue, 31 Aug 2021 09:56:34 +0000 (10:56 +0100)]
libgcc: Add missing runtime exception notices

Quoting from https://gcc.gnu.org/pipermail/gcc/2021-July/236716.html:

--------------------------------------------------------------------
It was pointed out to me off-list that config/aarch64/value-unwind.h
is missing the runtime exception.  It looks like a few other files
are too; a fuller list is:

libgcc/config/aarch64/value-unwind.h
libgcc/config/frv/frv-abi.h
libgcc/config/i386/value-unwind.h
libgcc/config/pa/pa64-hpux-lib.h

Certainly for the aarch64 file this was simply a mistake;
it seems to have been copied from the i386 version, both of which
reference the runtime exception but don't actually include it.
--------------------------------------------------------------------

Similarly, frv-abi.h referenced the exception but didn't include it.
pa64-hpux-lib.h was missing any reference to the exception.

The decision was that this was simply a mistake
[https://gcc.gnu.org/pipermail/gcc/2021-July/236717.html]:

--------------------------------------------------------------------
[…] It generally is
considered a textual omission.  The runtime library components of GCC
are intended to be licensed under the runtime exception, which was
granted and approved at the time of introduction.
--------------------------------------------------------------------

and that we should simply change all of the files above
[https://gcc.gnu.org/pipermail/gcc/2021-July/236719.html]:

--------------------------------------------------------------------
Please correct the text in the files. The files in libgcc used in the
GCC runtime are intended to be licensed with the runtime exception and
GCC previously was granted approval for that licensing and purpose.

[…]

The runtime exception explicitly was intended for this purpose and
usage at the time that GCC received approval to apply the exception.
--------------------------------------------------------------------

libgcc/
* config/aarch64/value-unwind.h: Add missing runtime exception
paragraph.
* config/frv/frv-abi.h: Likewise.
* config/i386/value-unwind.h: Likewise.
* config/pa/pa64-hpux-lib.h: Likewise.

3 years agomiddle-end/102129 - avoid TER of possibly trapping expressions
Richard Biener [Tue, 31 Aug 2021 07:13:20 +0000 (09:13 +0200)]
middle-end/102129 - avoid TER of possibly trapping expressions

The following avoids applying TER to possibly trapping expressions,
preventing a trapping FP multiplication to be moved across a call
that should not be executed.

2021-08-31  Richard Biener  <rguenther@suse.de>

PR middle-end/102129
* tree-ssa-ter.c (find_replaceable_in_bb): Do not move
possibly trapping expressions across calls.

3 years agogdb: Add a dependency between gdb and libbacktrace
Andrew Burgess [Mon, 30 Aug 2021 20:16:50 +0000 (21:16 +0100)]
gdb: Add a dependency between gdb and libbacktrace

GDB is going to start using libbacktrace, so add a build dependency
between the two modules.  This change needs to be added into the GCC
toplevel files, and then back-ported to the binutils-gdb repository.

2021-08-31  Andrew Burgess  <andrew.burgess@embecosm.com>

ChangeLog:

* Makefile.def: Add all-gdb dependency on all-libbacktrace.
* Makefile.in: Regenerate.

3 years agotree-ssa-ccp: Fix up bit_value_binop on RSHIFT_EXPR [PR102134]
Jakub Jelinek [Tue, 31 Aug 2021 08:29:23 +0000 (10:29 +0200)]
tree-ssa-ccp: Fix up bit_value_binop on RSHIFT_EXPR [PR102134]

As mentioned in the PR, this hunk is guarded with !wi::neg_p (r1val | r1mask, sgn)
which means if sgn is UNSIGNED, it is always true, but r1val | r1mask in
widest_int is still sign-extended.  That means wi::clz (arg) returns 0,
wi::get_precision (arg) returns some very large number
(WIDE_INT_MAX_PRECISION, on x86_64 576 bits) and width is 64, so we end up
with lzcount of -512 where the code afterwards expects a non-negative
lzcount.  For arg without the sign bit set the code works right, those
numbers are zero extended and so wi::clz must return wi::get_precision (arg) - width
plus number of leading zero bits within the width precision.
The patch fixes it by handling the sign-extension specially, either it could
be done through wi::neg_p (arg) check, but lzcount == 0 works identically.

2021-08-31  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/102134
* tree-ssa-ccp.c (bit_value_binop) <case RSHIFT_EXPR>: If sgn is
UNSIGNED and r1val | r1mask has MSB set, ensure lzcount doesn't
become negative.

* gcc.c-torture/execute/pr102134.c: New test.

3 years agoarm: Adjust test for soft fp targets.
prathamesh.kulkarni [Tue, 31 Aug 2021 07:25:19 +0000 (12:55 +0530)]
arm: Adjust test for soft fp targets.

gcc/testsuite/ChangeLog:
* gcc.target/arm/pr51534.c: Adjust test-case for soft fp targets.