platform/upstream/gcc.git
2 years agoaix: 64 bit AIX TLS libpthread dependency.
Clément Chigot [Thu, 12 Aug 2021 17:17:15 +0000 (13:17 -0400)]
aix: 64 bit AIX TLS libpthread dependency.

64bit XCOFF files will generated TLS access, with local-exec or
global-exec models, by an access to R13. Thus, there isn't
any reference to a TLS symbol.

The problem is that it allows programs with TLS to be compiled and
linked even without -pthread. Most of the time, it will result in
a segfault when trying to access a TLS variable. But sometimes, it
might create a memory corruption.

This patch forces a reference to __tls_get_addr() to ensure link
will fail without -pthread.

gcc/ChangeLog:
2021-08-11  Clément Chigot  <clement.chigot@atos.net>

* config/rs6000/rs6000.c (xcoff_tls_exec_model_detected): New.
(rs6000_legitimize_tls_address_aix): Use it.
(rs6000_xcoff_file_end): Add ".ref __tls_get_addr" when
xcoff_tls_exec_model_detected is true.

2 years agoImprove many SImode shifts on the H8/300H
Jeff Law [Sun, 15 Aug 2021 04:13:23 +0000 (00:13 -0400)]
Improve many SImode shifts on the H8/300H

As I've mentioned before, the H8/300H can only shift a single bit position at a time.  Naturally this means many shifts are implemented as loops.  There's a variety of special cases that we can do without loops by using rotates, sub-word moves, etc.  The general guidance for the port has been to only use inline or special sequences if they're shorter or just one instruction longer than the loop.

This was pretty reasonable guidance for QI/HI mode.  It was relaxed a bit about 10 years ago for HImode in particular where the kpit team realized they could save 50-100 cycles for some shifts by allowing 2 instructions of code growth over the loop implementation.

But they only re-tuned HImode shifts.  There's even bigger benefits for re-tuning SImode shifts.  There's cases where we can save close to 200 cycles by allowing 2 additional instructions.

This patch re-tunes SImode shifts on the H8/300H primarily by inlining more often or using a special sequence + inlining for residuals.  Both cases were already supported and this just uses those existing capabilities more often, so it was trivial to implement.  I think there's some cases were entirely new special sequences could be used, but I haven't tried those yet.

gcc/

* config/h8300/h8300.c (shift_alg_si): Retune H8/300H shifts
to allow a bit more code growth, saving many dozens of cycles.
(h8300_option_override): Adjus shift_alg_si if optimizing for
code size.
(get_shift_alg): Use special + inline shifts for residuals
in more cases.

2 years agolibgo: various fixes for Solaris support
Ian Lance Taylor [Sat, 14 Aug 2021 00:21:54 +0000 (17:21 -0700)]
libgo: various fixes for Solaris support

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/342189

2 years agoDaily bump.
GCC Administrator [Sun, 15 Aug 2021 00:16:27 +0000 (00:16 +0000)]
Daily bump.

2 years agoor1k: Use cmodel=large when building crtstuff
Stafford Horne [Tue, 20 Apr 2021 20:33:15 +0000 (05:33 +0900)]
or1k: Use cmodel=large when building crtstuff

When linking gcc runtime objects into large binaries the link may fail
with the below errors.  This will happen even if we are building with
-mcmodel=large.

    /home/shorne/work/openrisc/output/host/lib/gcc/or1k-buildroot-linux-uclibc/10.3.0/crtbeginS.o: in function `deregister_tm_clones':
    crtstuff.c:(.text+0x3c): relocation truncated to fit: R_OR1K_GOT16 against undefined symbol `_ITM_deregisterTMCloneTable'
    /home/shorne/work/openrisc/output/host/lib/gcc/or1k-buildroot-linux-uclibc/10.3.0/crtbeginS.o: in function `register_tm_clones':
    crtstuff.c:(.text+0xc0): relocation truncated to fit: R_OR1K_GOT16 against undefined symbol `_ITM_registerTMCloneTable'

This patch builds the gcc crtstuff binaries always with the
-mcmodel=large option to ensure they can be linked into large binaries.

libgcc/ChangeLog:

PR target/99783
* config.host (or1k-*, tmake_file): Add or1k/t-crtstuff.
* config/or1k/t-crtstuff: New file.

2 years agoor1k: Add mcmodel option to handle large GOTs
Stafford Horne [Mon, 5 Apr 2021 20:47:17 +0000 (05:47 +0900)]
or1k: Add mcmodel option to handle large GOTs

When building libgeos we get an error with:

    linux-uclibc/9.3.0/crtbeginS.o: in function `__do_global_dtors_aux':
    crtstuff.c:(.text+0x118): relocation truncated to fit: R_OR1K_GOT16 against symbol `__cxa_finalize' defined in .text section in
    /home/shorne/work/openrisc/3eb9f9d0f6d8274b2d19753c006bd83f7d536e3c/output/host/or1k-buildroot-linux-uclibc/sysroot/lib/libc.so.

This is caused by GOT code having a limit of 64k.  In OpenRISC this
looks to be the only relocation code pattern to be limited to 64k.

This patch allows specifying a new option -mcmodel=large which can be
used to generate 2 more instructions to construct 32-bit addresses for
up to 4G GOTs.

gcc/ChangeLog:

PR target/99783
* config/or1k/or1k-opts.h: New file.
* config/or1k/or1k.c (or1k_legitimize_address_1, print_reloc):
Support generating gotha relocations if -mcmodel=large is
specified.
* config/or1k/or1k.h (TARGET_CMODEL_SMALL, TARGET_CMODEL_LARGE):
New macros.
* config/or1k/or1k.opt (mcmodel=): New option.
* doc/invoke.texi (OpenRISC Options): Document mcmodel.

2 years agoDiagnose mismatches between array and scalar new and delete [PR101791].
Martin Sebor [Sat, 14 Aug 2021 19:25:41 +0000 (13:25 -0600)]
Diagnose mismatches between array and scalar new and delete [PR101791].

Resolves:
PR middle-end/101791 - missing warning on a mismatch between scalar and array forms of new and delete

gcc/ChangeLog:

PR middle-end/101791
* gimple-ssa-warn-access.cc (new_delete_mismatch_p): Use new argument
to valid_new_delete_pair_p.
* tree.c (valid_new_delete_pair_p): Add argument.
* tree.h (valid_new_delete_pair_p): Same.

gcc/testsuite/ChangeLog:

PR middle-end/101791
* g++.dg/warn/Wmismatched-new-delete-6.C: New test.
* g++.dg/warn/Wmismatched-new-delete-7.C: New test.

2 years agoi386: Fix ICE with V64QImode broadcast permutation with -mavx512f -mno-avx512bw
Jakub Jelinek [Sat, 14 Aug 2021 09:44:46 +0000 (11:44 +0200)]
i386: Fix ICE with V64QImode broadcast permutation with -mavx512f -mno-avx512bw

The testcase shows another problem, for TARGET_AVX512BW we have a single insn
doing broadcast from the first element, but don't have one for broadcast
of 2nd+ element (so for d->perm[0] we must return false), but for
TARGET_AVX512F && !TARGET_AVX512BW we don't even have support for that other
broadcast.  V64QImode case was just added to the AVX2 cases which had
gcc_assert (!TARGET_AVX2 || d->perm[0]);
but for V64QImode we actually need
gcc_assert (!TARGET_AVX512BW || d->perm[0]);

2021-08-14  Jakub Jelinek  <jakub@redhat.com>

PR target/101896
* config/i386/i386-expand.c (expand_vec_perm_broadcast_1)
<case E_V64QImode>: For this mode assert
!TARGET_AVX512BW || d->perm[0] rather than !TARGET_AVX2 || d->perm[0].

* gcc.target/i386/avx512f-pr101896.c: New test.

2 years agoDaily bump.
GCC Administrator [Sat, 14 Aug 2021 00:16:29 +0000 (00:16 +0000)]
Daily bump.

2 years agoFix xxeval predicates (PR 99921).
Michael Meissner [Fri, 13 Aug 2021 23:43:27 +0000 (19:43 -0400)]
Fix xxeval predicates (PR 99921).

I noticed that the xxeval built-in function used the altivec_register_operand
predicate.  Since it takes vsx registers, this might force the register
allocate to issue a move when it could use a traditional floating point
register.  This patch fixes that.

2021-08-13  Michael Meissner  <meissner@linux.ibm.com>

gcc/
PR target/99921
* config/rs6000/altivec.md (xxeval): Use register_predicate
instead of altivec_register_predicate.

2 years agoAdjust 'libgomp.oacc-c-c++-common/static-variable-1.c'
Thomas Schwinge [Fri, 13 Aug 2021 08:23:30 +0000 (10:23 +0200)]
Adjust 'libgomp.oacc-c-c++-common/static-variable-1.c'

... for 'gcc/gimplify.c:gimplify_scan_omp_clauses' changes in recent
commit d0befed793b94f3f407be44e6f69f81a02f5f073 "openmp: Add support
for OpenMP 5.1 masked construct".

libgomp/
* testsuite/libgomp.oacc-c-c++-common/static-variable-1.c: Adjust.

2 years agoWarn for reads from write-only arguments [PR101734].
Martin Sebor [Fri, 13 Aug 2021 18:21:20 +0000 (12:21 -0600)]
Warn for reads from write-only arguments [PR101734].

Resolves:
PR middle-end/101734 - missing warning reading from a write-only object

gcc/ChangeLog:

PR middle-end/101734
* tree-ssa-uninit.c (maybe_warn_read_write_only): New function.
(maybe_warn_operand): Call it.

gcc/testsuite/ChangeLog:

PR middle-end/101734
* gcc.dg/uninit-42.c: New test.

2 years agoruntime: use C cast syntax in stack.c
Ian Lance Taylor [Fri, 13 Aug 2021 17:15:45 +0000 (10:15 -0700)]
runtime: use C cast syntax in stack.c

Didn't notice earlier because this code is only used on systems that
do not support -fsplit-stack.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/342051

2 years agoipa: "naked" attribute implies "noipa" attribute
Martin Liska [Thu, 12 Aug 2021 15:26:51 +0000 (17:26 +0200)]
ipa: "naked" attribute implies "noipa" attribute

PR ipa/101354

gcc/ChangeLog:

* attribs.c (decl_attributes): Make naked functions "noipa"
  functions.

2 years agoipa: do not make localaliases for target_clones [PR101261]
Martin Liska [Fri, 13 Aug 2021 09:10:56 +0000 (11:10 +0200)]
ipa: do not make localaliases for target_clones [PR101261]

PR ipa/101261

gcc/ChangeLog:

* symtab.c (symtab_node::noninterposable_alias): Do not create
  local aliases for target_clone functions as the clonning pass
  rejects aliases.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr101261.c: New test.

2 years agoopts: do not repeat a string in errors
Martin Liska [Fri, 13 Aug 2021 09:18:40 +0000 (11:18 +0200)]
opts: do not repeat a string in errors

gcc/ChangeLog:

* opts.c (LIVE_PATCHING_OPTION): Define.
(control_options_for_live_patching): Use it in error messages.

2 years agoIntroduce EAF_NOREAD and cleanup EAF_UNUSED + ipa-modref
Jan Hubicka [Fri, 13 Aug 2021 08:04:52 +0000 (10:04 +0200)]
Introduce EAF_NOREAD and cleanup EAF_UNUSED + ipa-modref

this patch add EAF_NOREAD (as disucssed on IRC already) and fixes meaning
of EAF_UNUSED to be really unused and not "does not escape, is not clobbered,
read or returned" since we have separate flags for each of the properties
now.

Since number of flags has grown I refactored the code a bit to avoid
repeated uses of complex flag combinations and also simplified the logic of
merging.

Merging is bit tricky since we have flags that implies other flags
(like NOESCAPE implies NODIRECTESCAPE) but code that sets only NOESCAPE.
Perhaps it would make sense to update fnspecs to always set flag along with
all the implications, but for now I am handlingit in merge.

I made only trivial update to tree-ssa-structalias to handle EAF_NORETURN in
normal function handling, but not in pure functions.  The problem is that the
way constraints are generated for pure functions makes this difficult.
I think logical step is to track whether function reads/stores global memory
and rewrite the constraint generation so we can handle normal, pure and const
in unified manner.

Bootstrapped/regtested x86_64-linux, plan to commit it after furhter testing.

The patch improves alias oracle stats for cc1plus somewhat.

From:

Alias oracle query stats:
  refs_may_alias_p: 72380497 disambiguations, 82649832 queries
  ref_maybe_used_by_call_p: 495184 disambiguations, 73366950 queries
  call_may_clobber_ref_p: 259312 disambiguations, 263253 queries
  nonoverlapping_component_refs_p: 0 disambiguations, 38006 queries
  nonoverlapping_refs_since_match_p: 21157 disambiguations, 65698 must overlaps, 87756 queries
  aliasing_component_refs_p: 63141 disambiguations, 2164695 queries
  TBAA oracle: 25975753 disambiguations 61449632 queries
               12138220 are in alias set 0
               11316663 queries asked about the same object
               144 queries asked about the same alias set
               0 access volatile
               10472885 are dependent in the DAG
               1545967 are aritificially in conflict with void *

Modref stats:
  modref use: 23857 disambiguations, 754515 queries
  modref clobber: 1392162 disambiguations, 17753512 queries
  3450241 tbaa queries (0.194341 per modref query)
  534816 base compares (0.030125 per modref query)

PTA query stats:
  pt_solution_includes: 12394915 disambiguations, 20235925 queries
  pt_solutions_intersect: 1365299 disambiguations, 14638068 queries

To:

Alias oracle query stats:
  refs_may_alias_p: 72629640 disambiguations, 82903333 queries
  ref_maybe_used_by_call_p: 502474 disambiguations, 73612186 queries
  call_may_clobber_ref_p: 261806 disambiguations, 265659 queries
  nonoverlapping_component_refs_p: 0 disambiguations, 38007 queries
  nonoverlapping_refs_since_match_p: 21139 disambiguations, 65772 must overlaps, 87816 queries
  aliasing_component_refs_p: 63144 disambiguations, 2164330 queries
  TBAA oracle: 26059018 disambiguations 61571714 queries
               12158033 are in alias set 0
               11326115 queries asked about the same object
               144 queries asked about the same alias set
               0 access volatile
               10484493 are dependent in the DAG
               1543911 are aritificially in conflict with void *

Modref stats:
  modref use: 24008 disambiguations, 712712 queries
  modref clobber: 1395917 disambiguations, 17163694 queries
  3465657 tbaa queries (0.201918 per modref query)
  537591 base compares (0.031321 per modref query)

PTA query stats:
  pt_solution_includes: 12468934 disambiguations, 20295402 queries
  pt_solutions_intersect: 1391917 disambiguations, 14665265 queries

I think it is mostly due to better heandling of EAF_NODIRECTESCAPE.

Honza

gcc/ChangeLog:

2021-08-12  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref.c (dump_eaf_flags): Dump EAF_NOREAD.
(implicit_const_eaf_flags, implicit_pure_eaf_flags,
 ignore_stores_eaf_flags): New constants.
(remove_useless_eaf_flags): New function.
(eaf_flags_useful_p): Use it.
(deref_flags): Add EAF_NOT_RETURNED if flag is unused;
handle EAF_NOREAD.
(modref_lattice::init): Add EAF_NOREAD.
(modref_lattice::add_escape_point): Do not reacord escape point if
result is unused.
(modref_lattice::merge): EAF_NOESCAPE implies EAF_NODIRECTESCAPE;
use remove_useless_eaf_flags.
(modref_lattice::merge_deref): Use ignore_stores_eaf_flags.
(modref_lattice::merge_direct_load): Add EAF_NOREAD
(analyze_ssa_name_flags): Fix handling EAF_NOT_RETURNED
(analyze_parms): Use remove_useless_eaf_flags.
(ipa_merge_modref_summary_after_inlining): Use ignore_stores_eaf_flags.
(modref_merge_call_site_flags): Add caller and ecf_flags parameter;
use remove_useless_eaf_flags.
(modref_propagate_flags_in_scc): Update.
* ipa-modref.h: Turn eaf_flags_t back to char.
* tree-core.h (EAF_NOT_RETURNED): Fix.
(EAF_NOREAD): New constant
* tree-ssa-alias.c: (ref_maybe_used_by_call_p_1): Check for
EAF_NOREAD.
* tree-ssa-structalias.c (handle_rhs_call): Handle new flags.
(handle_pure_call): Likewise.

gcc/testsuite/ChangeLog:

2021-08-12  Jan Hubicka  <hubicka@ucw.cz>

* gcc.dg/tree-ssa/modref-6.c: Update.

2 years agoarm: Add check for arm_softfp_ok in pr98435.c.
prathamesh.kulkarni [Fri, 13 Aug 2021 06:58:50 +0000 (12:28 +0530)]
arm: Add check for arm_softfp_ok in pr98435.c.

gcc/testsuite/ChangeLog:
* gcc.target/arm/simd/pr98435.c: Add dg-require-effective-target
arm_softfp_ok.

2 years agolibbacktrace: fix b2test_buildid test on non-english locales
Sergei Trofimovich [Thu, 12 Aug 2021 22:09:08 +0000 (23:09 +0100)]
libbacktrace: fix b2test_buildid test on non-english locales

On LANG=ru_RU.UTF-8 'b2test_buildid' test fails due to localized readelf
output:

$ LANG=ru_RU.UTF-8 readelf -n b2test | fgrep 4e37e8f
    ID сборки: 4e37e8fead8d6e8b0a9dc95ea25cd784dff3a393
$ LANG=C readelf -n b2test | fgrep 4e37e8f
    Build ID: 4e37e8fead8d6e8b0a9dc95ea25cd784dff3a393

libbacktrace/

* install-debuginfo-for-buildid.sh.in: Force non-localized readelf
output with LANG=C.

2 years agolibgo: update to Go1.17rc2
Ian Lance Taylor [Fri, 30 Jul 2021 21:28:58 +0000 (14:28 -0700)]
libgo: update to Go1.17rc2

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/341629

2 years agoDaily bump.
GCC Administrator [Fri, 13 Aug 2021 00:16:43 +0000 (00:16 +0000)]
Daily bump.

2 years agoruntime: run gofmt on testdata/testwinsignal/main.go
Ian Lance Taylor [Fri, 13 Aug 2021 00:15:16 +0000 (17:15 -0700)]
runtime: run gofmt on testdata/testwinsignal/main.go

2 years agocompiler: store pointers to go:notinheap types indirectly
Ian Lance Taylor [Sat, 7 Aug 2021 01:04:43 +0000 (18:04 -0700)]
compiler: store pointers to go:notinheap types indirectly

This is the gofrontend version of https://golang.org/cl/264480.

For golang/go#42076

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/340609

2 years agoopenmp: Add support for OpenMP 5.1 masked construct
Jakub Jelinek [Thu, 12 Aug 2021 20:41:17 +0000 (22:41 +0200)]
openmp: Add support for OpenMP 5.1 masked construct

This construct has been introduced as a replacement for master
construct, but unlike that construct is slightly more general,
has an optional clause which allows to choose which thread
will be the one running the region, it can be some other thread
than the master (primary) thread with number 0, or it could be no
threads or multiple threads (then of course one needs to be careful
about data races).

It is way too early to deprecate the master construct though, we don't
even have OpenMP 5.0 fully implemented, it has been deprecated in 5.1,
will be also in 5.2 and removed in 6.0.  But even then it will likely
be a good idea to just -Wdeprecated warn about it and still accept it.

The patch also contains something I should have done much earlier,
for clauses that accept some integral expression where we only care
about the value, forces during gimplification that value into
either a min invariant (as before), SSA_NAME or a fresh temporary,
but never e.g. a user VAR_DECL, so that for those clauses we don't
need to worry about adjusting it.

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

gcc/
* tree.def (OMP_MASKED): New tree code.
* tree-core.h (enum omp_clause_code): Add OMP_CLAUSE_FILTER.
* tree.h (OMP_MASKED_BODY, OMP_MASKED_CLAUSES, OMP_MASKED_COMBINED,
OMP_CLAUSE_FILTER_EXPR): Define.
* tree.c (omp_clause_num_ops): Add OMP_CLAUSE_FILTER entry.
(omp_clause_code_name): Likewise.
(walk_tree_1): Handle OMP_CLAUSE_FILTER.
* tree-nested.c (convert_nonlocal_omp_clauses,
convert_local_omp_clauses): Handle OMP_CLAUSE_FILTER.
(convert_nonlocal_reference_stmt, convert_local_reference_stmt,
convert_gimple_call): Handle GIMPLE_OMP_MASTER.
* tree-pretty-print.c (dump_omp_clause): Handle OMP_CLAUSE_FILTER.
(dump_generic_node): Handle OMP_MASTER.
* gimple.def (GIMPLE_OMP_MASKED): New gimple code.
* gimple.c (gimple_build_omp_masked): New function.
(gimple_copy): Handle GIMPLE_OMP_MASKED.
* gimple.h (gimple_build_omp_masked): Declare.
(gimple_has_substatements): Handle GIMPLE_OMP_MASKED.
(gimple_omp_masked_clauses, gimple_omp_masked_clauses_ptr,
gimple_omp_masked_set_clauses): New inline functions.
(CASE_GIMPLE_OMP): Add GIMPLE_OMP_MASKED.
* gimple-pretty-print.c (dump_gimple_omp_masked): New function.
(pp_gimple_stmt_1): Handle GIMPLE_OMP_MASKED.
* gimple-walk.c (walk_gimple_stmt): Likewise.
* gimple-low.c (lower_stmt): Likewise.
* gimplify.c (is_gimple_stmt): Handle OMP_MASTER.
(gimplify_scan_omp_clauses): Handle OMP_CLAUSE_FILTER.  For clauses
that take one expression rather than decl or constant, force
gimplification of that into a SSA_NAME or temporary unless min
invariant.
(gimplify_adjust_omp_clauses): Handle OMP_CLAUSE_FILTER.
(gimplify_expr): Handle OMP_MASKED.
* tree-inline.c (remap_gimple_stmt): Handle GIMPLE_OMP_MASKED.
(estimate_num_insns): Likewise.
* omp-low.c (scan_sharing_clauses): Handle OMP_CLAUSE_FILTER.
(check_omp_nesting_restrictions): Handle GIMPLE_OMP_MASKED.  Adjust
diagnostics for existence of masked construct.
(scan_omp_1_stmt, lower_omp_master, lower_omp_1, diagnose_sb_1,
diagnose_sb_2): Handle GIMPLE_OMP_MASKED.
* omp-expand.c (expand_omp_synch, expand_omp, omp_make_gimple_edges):
Likewise.
gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_MASKED.
(enum pragma_omp_clause): Add PRAGMA_OMP_CLAUSE_FILTER.
* c-pragma.c (omp_pragmas_simd): Add masked construct.
* c-common.h (enum c_omp_clause_split): Add C_OMP_CLAUSE_SPLIT_MASKED
enumerator.
(c_finish_omp_masked): Declare.
* c-omp.c (c_finish_omp_masked): New function.
(c_omp_split_clauses): Handle combined masked constructs.
gcc/c/
* c-parser.c (c_parser_omp_clause_name): Parse filter clause name.
(c_parser_omp_clause_filter): New function.
(c_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_FILTER.
(OMP_MASKED_CLAUSE_MASK): Define.
(c_parser_omp_masked): New function.
(c_parser_omp_parallel): Handle parallel masked.
(c_parser_omp_construct): Handle PRAGMA_OMP_MASKED.
* c-typeck.c (c_finish_omp_clauses): Handle OMP_CLAUSE_FILTER.
gcc/cp/
* parser.c (cp_parser_omp_clause_name): Parse filter clause name.
(cp_parser_omp_clause_filter): New function.
(cp_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_FILTER.
(OMP_MASKED_CLAUSE_MASK): Define.
(cp_parser_omp_masked): New function.
(cp_parser_omp_parallel): Handle parallel masked.
(cp_parser_omp_construct, cp_parser_pragma): Handle PRAGMA_OMP_MASKED.
* semantics.c (finish_omp_clauses): Handle OMP_CLAUSE_FILTER.
* pt.c (tsubst_omp_clauses): Likewise.
(tsubst_expr): Handle OMP_MASKED.
gcc/testsuite/
* c-c++-common/gomp/clauses-1.c (bar): Add tests for combined masked
constructs with clauses.
* c-c++-common/gomp/clauses-5.c (foo): Add testcase for filter clause.
* c-c++-common/gomp/clause-dups-1.c (f1): Likewise.
* c-c++-common/gomp/masked-1.c: New test.
* c-c++-common/gomp/masked-2.c: New test.
* c-c++-common/gomp/masked-combined-1.c: New test.
* c-c++-common/gomp/masked-combined-2.c: New test.
* c-c++-common/goacc/uninit-if-clause.c: Remove xfails.
* g++.dg/gomp/block-11.C: New test.
* g++.dg/gomp/tpl-masked-1.C: New test.
* g++.dg/gomp/attrs-1.C (bar): Add tests for masked construct and
combined masked constructs with clauses in attribute syntax.
* g++.dg/gomp/attrs-2.C (bar): Likewise.
* gcc.dg/gomp/nesting-1.c (f1, f2): Add tests for masked construct
nesting.
* gfortran.dg/goacc/host_data-tree.f95: Allow also SSA_NAMEs in if
clause.
* gfortran.dg/goacc/kernels-tree.f95: Likewise.
libgomp/
* testsuite/libgomp.c-c++-common/masked-1.c: New test.

2 years agolibcpp: Fix ICE with -Wtraditional preprocessing [PR101638]
Jakub Jelinek [Thu, 12 Aug 2021 20:38:18 +0000 (22:38 +0200)]
libcpp: Fix ICE with -Wtraditional preprocessing [PR101638]

The following testcase ICEs in cpp_sys_macro_p, because cpp_sys_macro_p
is called for a builtin macro which doesn't use node->value.macro union
member but a different one and so dereferencing it ICEs.
As the testcase is distilled from contemporary glibc headers, it means
basically -Wtraditional now ICEs on almost everything.

The fix can be either the patch below, return true for builtin macros,
or we could instead return false for builtin macros, or the fix could
be also (untested):
--- libcpp/expr.c       2021-05-07 10:34:46.345122608 +0200
+++ libcpp/expr.c       2021-08-12 09:54:01.837556365 +0200
@@ -783,13 +783,13 @@ cpp_classify_number (cpp_reader *pfile,

       /* Traditional C only accepted the 'L' suffix.
          Suppress warning about 'LL' with -Wno-long-long.  */
-      if (CPP_WTRADITIONAL (pfile) && ! cpp_sys_macro_p (pfile))
+      if (CPP_WTRADITIONAL (pfile))
        {
          int u_or_i = (result & (CPP_N_UNSIGNED|CPP_N_IMAGINARY));
          int large = (result & CPP_N_WIDTH) == CPP_N_LARGE
                       && CPP_OPTION (pfile, cpp_warn_long_long);

-         if (u_or_i || large)
+         if ((u_or_i || large) && ! cpp_sys_macro_p (pfile))
            cpp_warning_with_line (pfile, large ? CPP_W_LONG_LONG : CPP_W_TRADITIONAL,
                                   virtual_location, 0,
                                   "traditional C rejects the \"%.*s\" suffix",
The builtin macros at least currently don't add any suffixes
or numbers -Wtraditional would like to warn about.  For floating
point suffixes, -Wtraditional calls cpp_sys_macro_p only right
away before emitting the warning, but in the above case the ICE
is because cpp_sys_macro_p is called even if the number doesn't
have any suffixes (that is I think always for builtin macros
right now).

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

PR preprocessor/101638
* macro.c (cpp_sys_macro_p): Return true instead of
crashing on builtin macros.

* gcc.dg/cpp/pr101638.c: New test.

2 years agoFix typo in fold-vec-load-builtin_vec_xl-* tests.
Michael Meissner [Thu, 12 Aug 2021 20:36:49 +0000 (16:36 -0400)]
Fix typo in fold-vec-load-builtin_vec_xl-* tests.

When I checked in the fix for running tests on power10 systems with
power10 code generation, I had a typo in the
fold-vec-load-builtin_vec_xl-* tests, swapping 'x' and 'v' in the p?lxv
pattern.

gcc/testsuite/
2021-08-12  Michael Meissner  <meissner@linux.ibm.com>

* gcc.target/powerpc/fold-vec-load-builtin_vec_xl-char.c: Fix
typo in regular expression.
* gcc.target/powerpc/fold-vec-load-builtin_vec_xl-double.c:
Likewise.
* gcc.target/powerpc/fold-vec-load-builtin_vec_xl-float.c:
Likewise.
* gcc.target/powerpc/fold-vec-load-builtin_vec_xl-int.c:
Likewise.
* gcc.target/powerpc/fold-vec-load-builtin_vec_xl-longlong.c:
Likewise.
* gcc.target/powerpc/fold-vec-load-builtin_vec_xl-short.c:
Likewise.

2 years ago[i386] Introduce scalar version of avx512f_vmscalef.
Uros Bizjak [Thu, 12 Aug 2021 19:18:46 +0000 (21:18 +0200)]
[i386] Introduce scalar version of avx512f_vmscalef.

2021-08-12  Uroš Bizjak  <ubizjak@gmail.com>

gcc/
PR target/98309
* config/i386/i386.md (avx512f_scalef<mode>2): New insn pattern.
(ldexp<mode>3): Use avx512f_scalef<mode>2.
(UNSPEC_SCALEF): Move from ...
* config/i386/sse.md (UNSPEC_SCALEF): ... here.

2 years agoFix condition testing void functions in ipa-split.
Jan Hubicka [Thu, 12 Aug 2021 18:52:54 +0000 (20:52 +0200)]
Fix condition testing void functions in ipa-split.

gcc/ChangeLog:

2021-08-12  Jan Hubicka  <hubicka@ucw.cz>

* ipa-split.c (consider_split): Fix condition testing void functions.

2 years agolibstdc++: Add additional overload of std::lerp [PR101870]
Jonathan Wakely [Thu, 12 Aug 2021 16:35:25 +0000 (17:35 +0100)]
libstdc++: Add additional overload of std::lerp [PR101870]

The [cmath.syn] p1 wording about additional overloads sufficient to
handle any arithmetic types also applies to std::lerp. This adds a new
overload of std::lerp that does the required promotions to support
arguments of arbitrary arithmetic types.

A new __promoted_t alias template is added, which the C++17 function
templates std::hypot and std::lerp can use to avoid instantiating the
__promote_3 class template.

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

PR libstdc++/101870
* include/c_global/cmath (hypot): Use __promoted_t.
(lerp): Add new overload accepting any arithmetic types.
* include/ext/type_traits.h (__promoted_t): New alias template.
* testsuite/26_numerics/lerp.cc: Moved to...
* testsuite/26_numerics/lerp/1.cc: ...here.
* testsuite/26_numerics/lerp/constexpr.cc: New test.
* testsuite/26_numerics/lerp/version.cc: New test.

2 years agolibstdc++: Make some #error strings consistent with other tests
Jonathan Wakely [Thu, 12 Aug 2021 17:02:40 +0000 (18:02 +0100)]
libstdc++: Make some #error strings consistent with other tests

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

* testsuite/26_numerics/lerp.cc: Add header name to #error.
* testsuite/26_numerics/midpoint/integral.cc: Likewise.
* testsuite/26_numerics/midpoint/version.cc: New test.

2 years agolibstdc++: Add [[nodiscard]] to experimental::randint
Jonathan Wakely [Thu, 12 Aug 2021 17:05:24 +0000 (18:05 +0100)]
libstdc++: Add [[nodiscard]] to experimental::randint

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

* include/experimental/random (experimental::randint): Add
nodiscard attribute.

2 years agoc++: fix ptrmemfunc template instantiation [PR101219]
Sergei Trofimovich [Fri, 6 Aug 2021 15:14:16 +0000 (16:14 +0100)]
c++: fix ptrmemfunc template instantiation [PR101219]

r12-1804 ("cp: add support for per-location warning groups.") among other
things removed warning suppression from a few places including ptrmemfuncs.

This exposed a bug in warning detection code as a reference to missing
BINFO (it's intentionally missing for ptrmemfunc types):

    crash_signal
        gcc/toplev.c:328
    perform_or_defer_access_check(tree_node*, tree_node*, tree_node*, int, access_failure_info*)
        gcc/cp/semantics.c:490
    finish_non_static_data_member(tree_node*, tree_node*, tree_node*)
        gcc/cp/semantics.c:2208
    ...

The change special cases ptrmemfuncs in templace substitution by using
build_ptrmemfunc_access_expr() instead of finish_non_static_data_member().

gcc/cp/ChangeLog:

PR c++/101219
* pt.c (tsubst_copy_and_build): Use build_ptrmemfunc_access_expr
to construct ptrmemfunc expression instantiation.

gcc/testsuite/ChangeLog:

PR c++/101219
* g++.dg/warn/pr101219.C: New test.

2 years agoEvaluate type arguments of sizeof that are structs of variable size [PR101838]
Martin Uecker [Thu, 12 Aug 2021 18:32:16 +0000 (20:32 +0200)]
Evaluate type arguments of sizeof that are structs of variable size [PR101838]

Evaluate type arguments of sizeof for all types of variable size
and not just for VLAs. This fixes PR101838 and some issues related
to PR29970 where statement expressions need to be evaluated so that
the size is well defined.

2021-08-12  Martin Uecker  <muecker@gwdg.de>

gcc/c/
PR c/101838
PR c/29970
* c-typeck.c (c_expr_sizeof_type): Evaluate
size expressions for structs of variable size.

gcc/testsuite/
PR c/101838
* gcc.dg/vla-stexp-2.c: New test.

2 years agoRemove legacy back threader.
Aldy Hernandez [Thu, 5 Aug 2021 06:01:47 +0000 (08:01 +0200)]
Remove legacy back threader.

At this point I don't see any use for the legacy mode, which I had
originally left in place during the transition.

This patch removes the legacy back threader, and cleans up the code a
bit.  There are no functional changes to the non-legacy code.

Tested on x86-64 Linux.

gcc/ChangeLog:

* doc/invoke.texi: Remove docs for threader-mode param.
* flag-types.h (enum threader_mode): Remove.
* params.opt: Remove threader-mode param.
* tree-ssa-threadbackward.c (class back_threader): Remove
path_is_unreachable_p.
Make find_paths private.
Add maybe_thread and thread_through_all_blocks.
Remove reference marker for m_registry.
Remove reference marker for m_profit.
(back_threader::back_threader): Adjust for registry and profit not
being references.
(dump_path): Move down.
(debug): Move down.
(class thread_jumps): Remove.
(class back_threader_registry): Remove m_all_paths.
Remove destructor.
(thread_jumps::thread_through_all_blocks): Move to back_threader
class.
(fsm_find_thread_path): Remove
(back_threader::maybe_thread): New.
(back_threader::thread_through_all_blocks): Move from
thread_jumps.
(back_threader_registry::back_threader_registry): Remove
m_all_paths.
(back_threader_registry::~back_threader_registry): Remove.
(thread_jumps::find_taken_edge): Remove.
(thread_jumps::check_subpath_and_update_thread_path): Remove.
(thread_jumps::maybe_register_path): Remove.
(thread_jumps::handle_phi): Remove.
(handle_assignment_p): Remove.
(thread_jumps::handle_assignment): Remove.
(thread_jumps::fsm_find_control_statement_thread_paths): Remove.
(thread_jumps::find_jump_threads_backwards): Remove.
(thread_jumps::find_jump_threads_backwards_with_ranger): Remove.
(try_thread_blocks): Rename find_jump_threads_backwards to
maybe_thread.
(pass_early_thread_jumps::execute): Same.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/ssa-dom-thread-7.c: Remove call into the legacy
code and adjust for ranger threader.

2 years agolibstdc++: Add #error to some files that depend on a specific standard mode
Jonathan Wakely [Thu, 12 Aug 2021 12:33:43 +0000 (13:33 +0100)]
libstdc++: Add #error to some files that depend on a specific standard mode

Give more explicit errors if these files are not built with the correct
-std options.

libstdc++-v3/ChangeLog:

* src/c++98/locale_init.cc: Require C++11.
* src/c++98/localename.cc: Likewise.
* src/c++98/misc-inst.cc: Require C++98.

2 years agoOpenMP 5.1: Add proc-bind 'primary' support
Tobias Burnus [Thu, 12 Aug 2021 13:48:28 +0000 (15:48 +0200)]
OpenMP 5.1: Add proc-bind 'primary' support

In OpenMP 5.1 "master thread" was changed to "primary thread" and
the proc_bind clause and the OMP_PROC_BIND environment variable
now take 'primary' as argument as alias for 'master', while the
latter is deprecated.
This commit accepts 'primary' and adds the named constant
omp_proc_bind_primary and changes 'master thread' in the
documentation; however, given that not even OpenMP 5.0 is
fully supported, omp_display_env and the dumps currently
still output 'master' and there is no deprecation warning
when using the 'master' in the proc_bind clause.

gcc/c/ChangeLog:

* c-parser.c (c_parser_omp_clause_proc_bind): Accept
'primary' as alias for 'master'.

gcc/cp/ChangeLog:

* parser.c (cp_parser_omp_clause_proc_bind): Accept
'primary' as alias for 'master'.

gcc/fortran/ChangeLog:

* gfortran.h (gfc_omp_proc_bind_kind): Add OMP_PROC_BIND_PRIMARY.
* dump-parse-tree.c (show_omp_clauses): Add TODO comment to
change 'master' to 'primary' in proc_bind for OpenMP 5.1.
* intrinsic.texi (OMP_LIB): Mention OpenMP 5.1; add
omp_proc_bind_primary.
* openmp.c (gfc_match_omp_clauses): Accept
'primary' as alias for 'master'.
* trans-openmp.c (gfc_trans_omp_clauses): Handle
OMP_PROC_BIND_PRIMARY.

gcc/ChangeLog:

* tree-core.h (omp_clause_proc_bind_kind): Add
OMP_CLAUSE_PROC_BIND_PRIMARY.
* tree-pretty-print.c (dump_omp_clause): Add TODO comment to
change 'master' to 'primary' in proc_bind for OpenMP 5.1.

libgomp/ChangeLog:

* env.c (parse_bind_var): Accept 'primary' as alias for
'master'.
(omp_display_env): Add TODO comment to
change 'master' to 'primary' in proc_bind for OpenMP 5.1.
* libgomp.texi: Change 'master thread' to 'primary thread'
in line with OpenMP 5.1.
(omp_get_proc_bind): Add omp_proc_bind_primary and note that
omp_proc_bind_master is an alias of it.
(OMP_PROC_BIND): Mention 'PRIMARY'.
* omp.h.in (__GOMP_DEPRECATED_5_1): Define.
(omp_proc_bind_primary): Add.
(omp_proc_bind_master): Deprecate for OpenMP 5.1.
* omp_lib.f90.in (omp_proc_bind_primary): Add.
(omp_proc_bind_master): Deprecate for OpenMP 5.1.
* omp_lib.h.in (omp_proc_bind_primary): Add.
* testsuite/libgomp.c/affinity-1.c: Check that
'primary' works and is identical to 'master'.

gcc/testsuite/ChangeLog:

* c-c++-common/gomp/pr61486-2.c: Duplicate one proc_bind(master)
testcase and test proc_bind(primary) instead.
* gfortran.dg/gomp/affinity-1.f90: Likewise.

2 years agoarc: Small data doesn't need fcommon option
Claudiu Zissulescu [Thu, 12 Aug 2021 11:21:22 +0000 (14:21 +0300)]
arc: Small data doesn't need fcommon option

ARC backend is defaulting to -fcommon. This is not anylonger needed, remove it.

gcc/
2021-08-12  Claudiu Zissulescu  <claziss@synopsys.com>

* common/config/arc/arc-common.c (arc_option_init_struct): Remove
fno-common reference.
* config/arc/arc.c (arc_override_options): Remove overriding of
flag_no_common.

Signed-off-by: Claudiu Zissulescu <claziss@synopsys.com>
2 years agogfortran.dg/PR82376.f90: Avoid matching a file-path
Hans-Peter Nilsson [Wed, 11 Aug 2021 21:33:40 +0000 (23:33 +0200)]
gfortran.dg/PR82376.f90: Avoid matching a file-path

I had a file-path to sources with the substring "new" in it,
and (only) this test regressed compared to results from
another build without "new" in the name.

The test does
 ! { dg-final { scan-tree-dump-times "new" 4 "original" } }
i.e. the contents of the tree-dump-file .original needs to match
the undelimited string "new" exactly four times.  Very brittle.

In the dump-file, there are three lines with calls to new:
     D.908 = new ((integer(kind=4) *) data);
 integer(kind=4) * new (integer(kind=4) & data)
   static integer(kind=4) * new (integer(kind=4) &);

But, there's also a line, which for me and cris-elf looked like:
 _gfortran_runtime_error_at (&"At line 46 of file
  /X/xyzzynewfrob/gcc/testsuite/gfortran.dg/PR82376.f90"[1]{lb: 1 sz: 1},
  &"Pointer actual argument \'new\' is not associated"[1]{lb: 1 sz: 1});
The fourth match is obviously intended to match this line, but only
with *one* match, whereas the path can as above yield another hit.

With Tcl, the regexp for matching the " " *and* the "'"
*and* the "\" gets a bit unsightly, so I suggest just
matching the "new" calls, which according to the comment in
the test is the key point.  You can't have a file-path with
spaces and parentheses in a gcc build.  I'm also making use
of {} rather than "" needing one level of quoting; the "\("
is needed because the matched string is a regexp.

testsuite:
* gfortran.dg/PR82376.f90: Robustify match.

2 years agoi386: Fix up V32HImode permutations with -mno-avx512bw [PR101860]
Jakub Jelinek [Thu, 12 Aug 2021 09:26:57 +0000 (11:26 +0200)]
i386: Fix up V32HImode permutations with -mno-avx512bw [PR101860]

My patch from yesterday apparently broke some V32HImode permutations
as the testcase shows.
The first function assumed it would never be called in d->testing_p mode
and so went right away into emitting the code.
And the second one assumed V32HImode would never reach it, which now
can for the !TARGET_AVX512BW case.  We don't have a instruction
in that case though.

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

PR target/101860
* config/i386/i386-expand.c (ix86_expand_vec_one_operand_perm_avx512):
If d->testing_p, return true after performing checks instead of
actually expanding the insn.
(expand_vec_perm_broadcast_1): Handle V32HImode - assert
!TARGET_AVX512BW and return false.

* gcc.target/i386/avx512f-pr101860.c: New test.

2 years agoopenmp: Diagnose syntax mismatches between declare target and end declare target
Jakub Jelinek [Thu, 12 Aug 2021 07:26:27 +0000 (09:26 +0200)]
openmp: Diagnose syntax mismatches between declare target and end declare target

OpenMP 5.1 says:
For any directive that has a paired end directive, including those with a begin
and end pair, both directives must use either the attribute syntax or the
pragma syntax.

The following patch enforces it with the only pair so far recognized in C++
(Fortran has many, but on the other side doesn't have attribute syntax).

While I initially wanted to use vec<bool, va_gc> *member; in there, that
unfortunately doesn't work, one gets linker errors and I guess it is fixable,
but for begin declare target we'll need a struct anyway to store device_type
etc.

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

* cp-tree.h (omp_declare_target_attr): New type.
(struct saved_scope): Change type of omp_declare_target_attribute
from int to vec<omp_declare_target_attr, va_gc> * and move it.
* parser.c (cp_parser_omp_declare_target): Instead of
incrementing scope_chain->omp_declare_target_attribute, push
a struct containing parser->lexer->in_omp_attribute_pragma to
the vector.
(cp_parser_omp_end_declare_target): Instead of decrementing
scope_chain->omp_declare_target_attribute, pop a structure
from it.  Diagnose mismatching declare target vs.
end declare target syntax.
* semantics.c (finish_translation_unit): Use vec_safe_length
and vec_safe_truncate on scope_chain->omp_declare_target_attributes.
* decl2.c (cplus_decl_attributes): Use vec_safe_length
on scope_chain->omp_declare_target_attributes.

* g++.dg/gomp/attrs-12.C: New test.

2 years agoopenmp: Diagnose another case of mixing parameter and attribute syntax
Jakub Jelinek [Thu, 12 Aug 2021 07:18:23 +0000 (09:18 +0200)]
openmp: Diagnose another case of mixing parameter and attribute syntax

This patch diagnoses cases like:
  #pragma omp parallel
  [[omp::directive (declare simd)]] int foo ();
or
  #pragma omp taskgroup
  int bar [[omp::directive (declare simd)]] (int);
where the pragma is on the same declaration statement as the declare simd
attribute.

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

* parser.c (cp_parser_lambda_body): Add temp overrides
for parser->{omp_declare_simd,oacc_routine,omp_attrs_forbidden_p}.
(cp_parser_statement): Restore parser->omp_attrs_forbidden_p for
cp_parser_declaration_statement.
(cp_parser_default_argument): Add temp override for
parser->omp_attrs_forbidden_p.
(cp_parser_late_parsing_omp_declare_simd): Diagnose declare simd
or declare variant in attribute syntax on a declaration immediately
following an OpenMP construct in pragma syntax.

* g++.dg/gomp/attrs-11.C: Add new tests.

2 years agoc++: Fix ICE on defaulted spaceship with pointer return type [PR94162]
Jakub Jelinek [Thu, 12 Aug 2021 07:16:13 +0000 (09:16 +0200)]
c++: Fix ICE on defaulted spaceship with pointer return type [PR94162]

The spaceship-synth-neg6.C testcase ICEs because we call cat_tag_for
on the explicit return type, but pointer types don't have
TYPE_LINKAGE_IDENTIFIER.  The patch fixes that by checking for
CLASS_TYPE_P only and also adds verification that it is in std
namespace, so we don't return non-cc_last for my_namespace::partial_ordering.

The g++.dg/cpp2a/spaceship-synth11.C testcase is from a PR that has been
fixed with r12-619-gfc178519771db508c03611cff4a1466cf67fce1d (but
not backported to 11).

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

gcc/cp/
PR c++/94162
* method.c (cat_tag_for): Return cc_last for !CLASS_TYPE_P
or for classes not in std namespace.
gcc/testsuite/
PR c++/99429
* g++.dg/cpp2a/spaceship-synth11.C: New test.

PR c++/94162
* g++.dg/cpp2a/spaceship-synth-neg6.C: New test.

2 years agoopenmp: Diagnose omp::directive/sequence on using-directive
Jakub Jelinek [Thu, 12 Aug 2021 07:10:46 +0000 (09:10 +0200)]
openmp: Diagnose omp::directive/sequence on using-directive

With the using-directive parsing changes, we now emit only a warning
for [[omp::directive (...)]] on using-directive.  While that is right
without -fopenmp/-fopenmp-simd, when OpenMP is enabled, that should
be an error as OpenMP (is going to) disallow such attributes there
as they do not appertain to a statement.

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

* name-lookup.c (finish_using_directive): Diagnose omp::directive
or omp::sequence attributes on using-directive.

* g++.dg/gomp/attrs-11.C: Adjust expected diagnostics.

2 years agoc++: Fix up parsing of attributes for using-directive
Jakub Jelinek [Thu, 12 Aug 2021 07:09:39 +0000 (09:09 +0200)]
c++: Fix up parsing of attributes for using-directive

As I've said earlier and added xfails in gen-attrs-76.C test,
https://eel.is/c++draft/namespace.udir#nt:using-directive
has attribute-specifier-seq[opt] at the start, not at the end before ;
as gcc is expecting.
IMHO we should continue parsing at the end the GNU attributes
because using namespace N __attribute__((strong));, while not supported
anymore, used to be supported in the past, but my code searches for
using namespace N [[gnu::strong]]; didn't reveal anything at all.

2021-08-12  Jakub Jelinek  <jakub@redhat.com>

* parser.c (cp_parser_block_declaration): Call
cp_parser_using_directive for C++11 attributes followed by
using namespace tokens.
(cp_parser_using_directive): Parse C++11 attributes at the start
of the directive rather than at the end, only parse GNU attributes
at the end.

* g++.dg/lookup/strong-using.C: Add test using [[gnu::strong]]
as well.
* g++.dg/lookup/strong-using2.C: Likewise.
* g++.dg/cpp0x/gen-attrs-58.C: Move alignas(int) before
using namespace.
* g++.dg/cpp0x/gen-attrs-59.C: Move alignas(X) before
using namespace, add tests for alignas before semicolon.
* g++.dg/cpp0x/gen-attrs-76.C: Remove xfails.  Add test for
C++11 attributes on using directive before semicolon.

2 years agoMake -no-pie option work for native Windows
Eric Botcazou [Thu, 12 Aug 2021 07:30:31 +0000 (09:30 +0200)]
Make -no-pie option work for native Windows

Binutils 2.36/2.37 generate PIE executables by default on native Windows
(because --dynamicbase is the default) so it makes sense to have a simple
way to counter that and -no-pie seems appropriate, all the more so that
it is automatically passed when building the compiler itself.

gcc/
* configure.ac (PE linker --disable-dynamicbase support): New check.
* configure: Regenerate.
* config.in: Likewise.
* config/i386/mingw32.h (LINK_SPEC_DISABLE_DYNAMICBASE): New define.
(LINK_SPEC): Use it.
* config/i386/mingw-w64.h (LINK_SPEC_DISABLE_DYNAMICBASE): Likewise.
(LINK_SPEC): Likewise.

2 years agoCombine avx_vec_concatv16si and avx512f_zero_extendv16hiv16si2_1 to avx512f_zero_exte...
liuhongt [Wed, 11 Aug 2021 06:00:00 +0000 (14:00 +0800)]
Combine avx_vec_concatv16si and avx512f_zero_extendv16hiv16si2_1 to avx512f_zero_extendv16hiv16si2_2.

Add define_insn_and_split to combine avx_vec_concatv16si/2 and
avx512f_zero_extendv16hiv16si2_1 since the latter already zero_extend
the upper bits, similar for other patterns which are related to
pmovzx{bw,wd,dq}.

It will do optimization like

-       vmovdqa %ymm0, %ymm0    # 7     [c=4 l=6]  avx_vec_concatv16si/2
        vpmovzxwd       %ymm0, %zmm0    # 22    [c=4 l=6]  avx512f_zero_extendv16hiv16si2
        ret             # 25    [c=0 l=1]  simple_return_internal

gcc/ChangeLog:

PR target/101846
* config/i386/sse.md (*avx2_zero_extendv16qiv16hi2_2): New
post_reload define_insn_and_split.
(*avx512bw_zero_extendv32qiv32hi2_2): Ditto.
(*sse4_1_zero_extendv8qiv8hi2_4): Ditto.
(*avx512f_zero_extendv16hiv16si2_2): Ditto.
(*avx2_zero_extendv8hiv8si2_2): Ditto.
(*sse4_1_zero_extendv4hiv4si2_4): Ditto.
(*avx512f_zero_extendv8siv8di2_2): Ditto.
(*avx2_zero_extendv4siv4di2_2): Ditto.
(*sse4_1_zero_extendv2siv2di2_4): Ditto.
(VI248_256, VI248_512, VI148_512, VI148_256, VI148_128): New
mode iterator.

gcc/testsuite/ChangeLog:

PR target/101846
* gcc.target/i386/pr101846-1.c: New test.

2 years agoc++: constexpr std::construct_at on empty field [PR101663]
Patrick Palka [Thu, 12 Aug 2021 00:59:53 +0000 (20:59 -0400)]
c++: constexpr std::construct_at on empty field [PR101663]

Here during constexpr evaluation of

  std::construct_at(&a._M_value)

we find ourselves in cxx_eval_store_expression where the target object
is 'a._M_value' and the initializer is {}.  Since _M_value is an empty
[[no_unique_address]] member we don't create a sub-CONSTRUCTOR for it,
so we end up in the early exit code path for empty stores with mismatched
types and trip over the assert therein

  gcc_assert (is_empty_class (TREE_TYPE (init)) && !lval);

because lval is true.  The reason it's true is because the INIT_EXPR in
question is the LHS of a COMPOUND_EXPR, and evaluation of the LHS is
always performed with lval=true (to indicate there's no lvalue-to-rvalue
conversion).

This patch makes the code path in question handle the lval=true case
appropriately rather than asserting.  In passing, it also consolidates
the duplicate implementations of std::construct_at/destroy_at in some
of the C++20 constexpr tests into a common header file.

PR c++/101663

gcc/cp/ChangeLog:

* constexpr.c (cxx_eval_store_expression): Handle the lval=true
case in the early exit code path for empty stores with mismatched
types.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/construct_at.h: New convenience header file that
defines minimal implementations of std::construct_at/destroy_at,
split out from ...
* g++.dg/cpp2a/constexpr-new5.C: ... here.
* g++.dg/cpp2a/constexpr-new6.C: Use the header.
* g++.dg/cpp2a/constexpr-new14.C: Likewise.
* g++.dg/cpp2a/constexpr-new20.C: New test.

2 years agoDaily bump.
GCC Administrator [Thu, 12 Aug 2021 00:16:28 +0000 (00:16 +0000)]
Daily bump.

2 years agolibstdc++: Fix test that fails randomly [PR101866]
Jonathan Wakely [Wed, 11 Aug 2021 21:11:19 +0000 (22:11 +0100)]
libstdc++: Fix test that fails randomly [PR101866]

This test assumes that the same sequence of three values cannot occur,
which is incorect. It's unlikely, but not impossible.

Perform the check in a loop, so that in the unlikely event of an
identical sequence, we retry. If the library code is buggy it will keep
producing the same sequence and the test will time out. If the code is
working correctly then we will usually break out of the loop after one
iteration, or very rarely after two or three.

libstdc++-v3/ChangeLog:

PR libstdc++/101866
* testsuite/experimental/random/randint.cc: Loop and retry if
reseed() produces the same sequence.

2 years agors6000: Add available-everywhere and ancient builtins
Bill Schmidt [Wed, 11 Aug 2021 19:59:17 +0000 (14:59 -0500)]
rs6000: Add available-everywhere and ancient builtins

2021-08-11  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add always, power5, and
power6 stanzas.

2 years agors6000: Add VSX builtins
Bill Schmidt [Wed, 11 Aug 2021 19:56:26 +0000 (14:56 -0500)]
rs6000: Add VSX builtins

2021-08-11  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add vsx stanza.

2 years agors6000: Add the rest of the [altivec] stanza to the builtins file
Bill Schmidt [Thu, 10 Jun 2021 16:14:51 +0000 (11:14 -0500)]
rs6000: Add the rest of the [altivec] stanza to the builtins file

2021-06-10  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Finish altivec stanza.
* config/rs6000/rs6000-call.c (rs6000_init_builtins): Move
initialization of pcvoid_type_node here...
(altivec_init_builtins): ...from here.
* config/rs6000/rs6000.h (rs6000_builtin_type_index): Add
RS6000_BTI_const_ptr_void.
(pcvoid_type_node): New macro.

2 years agoc++: parameterized requires-expr as default argument [PR101725]
Patrick Palka [Wed, 11 Aug 2021 20:53:53 +0000 (16:53 -0400)]
c++: parameterized requires-expr as default argument [PR101725]

Here we're rejecting the default template argument

  requires (T t) { x(t); }

because we consider the 't' in the requirement to be a local variable
(according to local_variable_p), and we generally forbid local variables
from appearing inside default arguments.  We can perhaps fix this by
giving special treatment to parameters introduced by requires-expressions,
but DR 2082 relaxed the restriction about local variables appearing within
default arguments to permit them inside unevaluated operands thereof.
So this patch just implements DR 2082 which also fixes this PR since a
requires-expression is an unevaluated context.

PR c++/101725
DR 2082

gcc/cp/ChangeLog:

* cp-tree.h (unevaluated_p): Return true for REQUIRES_EXPR.
* decl.c (local_variable_p_walkfn): Don't walk into unevaluated
operands.
* parser.c (cp_parser_primary_expression) <case CPP_NAME>: Never
reject uses of local variables in unevaluated contexts.
* tree.c (cp_walk_subtrees) <case REQUIRES_EXPR>: Increment
cp_unevaluated_operand.  Use cp_walk_tree directly instead of
WALK_SUBTREE to avoid the goto.  Use REQUIRES_EXPR_REQS instead
of TREE_OPERAND directly.

gcc/testsuite/ChangeLog:

* g++.dg/DRs/dr2082.C: New test.
* g++.dg/cpp2a/concepts-uneval4.C: New test.

2 years agoc++: Optimize constinit thread_local vars [PR101786]
Jakub Jelinek [Wed, 11 Aug 2021 20:00:29 +0000 (22:00 +0200)]
c++: Optimize constinit thread_local vars [PR101786]

The paper that introduced constinit mentioned in rationale that constinit
can be used on externs as well and that it can be used to avoid the
thread_local initialization wrappers, because the standard requires that
if constinit is present on any declaration, it is also present on the
initialization declaration, even if it is in some other TU etc.

There is a small problem though, we use the tls wrappers not just if
the thread_local variable needs dynamic initialization, but also when
it has static initialization, but non-trivial destructor, as the
"dynamic initialization" in that case needs to register the destructor.

So, the following patch optimizes constinit thread_local vars only
if we can prove they will not have non-trivial destructors.  That includes
the case where we have incomplete type where we don't know and need to
conservatively assume the type will have non-trivial destructor at the
initializing declaration side.

2021-08-11  Jakub Jelinek  <jakub@redhat.com>

PR c++/101786
* decl2.c (var_defined_without_dynamic_init): Return true for
DECL_DECLARED_CONSTINIT_P with complete type and trivial destructor.

* g++.dg/cpp2a/constinit16.C: New test.

2 years agoc++: recognize class-scope non-template dguides [PR79501]
Patrick Palka [Wed, 11 Aug 2021 19:59:22 +0000 (15:59 -0400)]
c++: recognize class-scope non-template dguides [PR79501]

It looks like we still don't recognize class-scope non-template
deduction guides even after r12-2260.   This is because deduction guides
are tagged as such in cp_parser_init_declarator after calling
cp_parser_declarator, but in cp_parser_member_declaration we call
cp_parser_declarator directly.

So let's tag them in cp_parser_member_declaration as well.

PR c++/79501

gcc/cp/ChangeLog:

* parser.c (maybe_adjust_declarator_for_dguide): New, split
out from ...
(cp_parser_init_declarator): ... here.
(cp_parser_member_declaration): Use it.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1z/class-deduction98.C: New test.

2 years agoc++: most vexing parse and braced CTAD [PR89062]
Patrick Palka [Wed, 11 Aug 2021 19:58:30 +0000 (15:58 -0400)]
c++: most vexing parse and braced CTAD [PR89062]

Here grokdeclarator is emitting the error

  error: class template placeholder ‘Foo’ not permitted in this context

during the tentative (and ultimately futile) parse of 'x' as a function
declaration.  This happens because when parsing 'Foo{1}',
cp_parser_parameter_declaration yields a parameter declaration with no
declarator and whose type is a CTAD placeholder, and stops short of
consuming the '{'.  The caller cp_parser_parameter_declaration_list then
calls grokdeclarator on this declarator, hence the error, and soon
thereafter we abort this tentative parse since the next token '{'
doesn't make sense in the context of a parameter list.

Note that we don't have this issue with parenthesized CTAD

  Foo<int> x(Foo(1));

because in this case cp_parser_direct_declarator (called indirectly from
c_p_p_d) consumes the '(' and returns cp_error_declarator instead of a
NULL declarator (and also simulates a parse error), and grokdeclarator
exits early for this declarator without emitting any error.

Since grokdeclarator doesn't take a 'complain' parameter, to fix this we
need to avoid calling grokdeclarator in this situation.  To that end
this patch makes c_p_p_d simulate an error when a construct is a CTAD
expression and definitely not a parameter declaration, so that c_p_p_d_l
can avoid calling grokdeclarator by checking for this simulated error.
Alternatively we could keep all this logic inside c_p_p_d_l and not
touch c_p_p_d at all, but this approach seems slightly less adhoc.

PR c++/89062

gcc/cp/ChangeLog:

* parser.c (cp_parser_parameter_declaration_list): Don't call
grokdeclarator if cp_parser_error_occurred.
(cp_parser_parameter_declaration): Simulate an error if we see
the beginning of a CTAD form, i.e. if we see an opening brace
after the decl-specifier-seq and the type is a CTAD placeholder.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1z/class-deduction97.C: New test.

2 years agolibstdc++: Define std::is_pointer_interconvertible_base_of for C++20
Jonathan Wakely [Tue, 10 Aug 2021 14:37:23 +0000 (15:37 +0100)]
libstdc++: Define std::is_pointer_interconvertible_base_of for C++20

Implement these traits using the new built-ins that Jakub added
recently.

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

* include/std/type_traits (__cpp_lib_is_pointer_interconvertible)
(is_pointer_interconvertible_base_of_v)
(is_pointer_interconvertible_base_of): Define for C++20.
* include/std/version (__cpp_lib_is_pointer_interconvertible):
Define.
* testsuite/23_containers/span/layout_compat.cc: Use correct
feature test macro for std::is_layout_compatible_v.
* testsuite/20_util/is_pointer_interconvertible/value.cc: New test.
* testsuite/20_util/is_pointer_interconvertible/version.cc: New test.

2 years agolibstdc++: Fix missing descriptions in ChangeLog entries
Jonathan Wakely [Wed, 11 Aug 2021 13:05:01 +0000 (14:05 +0100)]
libstdc++: Fix missing descriptions in ChangeLog entries

2 years agoFortran: Fix c_float128 and c_float128_complex definitions.
Sandra Loosemore [Tue, 3 Aug 2021 23:21:16 +0000 (16:21 -0700)]
Fortran: Fix c_float128 and c_float128_complex definitions.

gfc_float128_type_node is only non-NULL on targets that support a
128-bit type that is not long double.  Use float128_type_node instead
when computing the value of the kind constants c_float128 and
c_float128_complex from the ISO_C_BINDING intrinsic module; this also
ensures it actually corresponds to __float128 (the IEEE encoding) and
not some other 128-bit floating-point type.

2021-08-11  Sandra Loosemore  <sandra@codesourcery.com>

gcc/fortran/
* iso-c-binding.def (c_float128, c_float128_complex): Check
float128_type_node instead of gfc_float128_type_node.
* trans-types.c (gfc_init_kinds, gfc_build_real_type):
Update comments re supported 128-bit floating-point types.

2 years agoFix gcc.dg/lto/pr48622_0.c testcase
Richard Biener [Wed, 11 Aug 2021 14:28:33 +0000 (16:28 +0200)]
Fix gcc.dg/lto/pr48622_0.c testcase

This fixes the testcase to not rely on the reference to
ashift_qi_1 being optimized out by RTL optimization via
help of the initregs pass that changes comparisons of
uninitialized data with a comparison that is always false.

2021-08-11  Richard Biener  <rguenther@suse.de>

* gcc.dg/lto/pr48622_1.c: Provide non-LTO definition
of ashift_qi_1.

2 years agotarget/101788 - avoid decomposing hard-register "loads"
Richard Biener [Wed, 11 Aug 2021 12:10:52 +0000 (14:10 +0200)]
target/101788 - avoid decomposing hard-register "loads"

This avoids decomposing hard-register accesses that masquerade as
loads.

2021-08-11  Richard Biener  <rguenther@suse.de>

PR target/101877
* tree-ssa-forwprop.c (pass_forwprop::execute): Do not decompose
hard-register accesses.

2 years agoAdjust volatile handling of the operand scanner
Richard Biener [Mon, 9 Aug 2021 11:12:08 +0000 (13:12 +0200)]
Adjust volatile handling of the operand scanner

The GIMPLE SSA operand scanner handles COMPONENT_REFs that are
not marked TREE_THIS_VOLATILE but have a TREE_THIS_VOLATILE
FIELD_DECL as volatile.  That's inconsistent in how TREE_THIS_VOLATILE
testing on GENERIC refs works which requires operand zero of
component references to mirror TREE_THIS_VOLATILE to the ref
so that testing TREE_THIS_VOLATILE on the outermost reference
is enough to determine the volatileness.

The following patch thus removes FIELD_DECL scanning from
the GIMPLE SSA operand scanner, possibly leaving fewer stmts
marked as gimple_has_volatile_ops.

It shows we miss at least one case in the fortran frontend, though
there's a suspicious amount of COMPONENT_REF creation compared
to little setting of TREE_THIS_VOLATILE.  This fixes the FAIL
of gfortran.dg/volatile11.f90 that would otherwise occur.

Visually inspecting fortran/ reveals a bunch of likely to fix
cases but I don't know the constraints of 'volatile' uses in
the fortran language to assess whether some of these are not
necessary.

2021-08-09  Richard Biener  <rguenther@suse.de>

gcc/
* tree-ssa-operands.c (operands_scanner::get_expr_operands):
Do not look at COMPONENT_REF FIELD_DECLs TREE_THIS_VOLATILE
to determine has_volatile_ops.

gcc/fortran/
* trans-common.c (create_common): Set TREE_THIS_VOLATILE on the
COMPONENT_REF if the field is volatile.

2 years agoSmall tweak to expand_used_vars
Eric Botcazou [Wed, 11 Aug 2021 13:31:32 +0000 (15:31 +0200)]
Small tweak to expand_used_vars

This completes the replacement of DECL_ATTRIBUTES (current_function_decl)
with the attribs local variable.

gcc/
* cfgexpand.c (expand_used_vars): Reuse attribs local variable.

2 years agoFix min_flags handling in mod-ref
Jan Hubicka [Wed, 11 Aug 2021 13:01:39 +0000 (15:01 +0200)]
Fix min_flags handling in mod-ref

gcc/ChangeLog:

2021-08-11  Jan Hubicka  <hubicka@ucw.cz>
    Alexandre Oliva  <oliva@adacore.com>

* ipa-modref.c (modref_lattice::dump): Fix escape_point's min_flags
dumping.
(modref_lattice::merge_deref): Fix handling of indirect scape points.
(update_escape_summary_1): Likewise.
(update_escape_summary): Likewise.
(ipa_merge_modref_summary_after_inlining): Likewise.

gcc/testsuite/ChangeLog:

* c-c++-common/modref-dse.c: New test.

2 years agomiddle-end/101858 - avoid shift of pointer in folding
Richard Biener [Wed, 11 Aug 2021 08:26:18 +0000 (10:26 +0200)]
middle-end/101858 - avoid shift of pointer in folding

This makes sure to not generate a shift of pointer types in
simplification of X < (cast) (1 << Y).

2021-08-11  Richard Biener  <rguenther@suse.de>

PR middle-end/101858
* fold-const.c (fold_binary_loc): Guard simplification
of  X < (cast) (1 << Y) to integer types.

* gcc.dg/pr101858.c: New testcase.

2 years agotree-optimization/101861 - fix gather use for non-gather refs
Richard Biener [Wed, 11 Aug 2021 09:49:10 +0000 (11:49 +0200)]
tree-optimization/101861 - fix gather use for non-gather refs

My previous change broke the usage of gather for strided loads.
The following fixes it.

2021-08-11  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101861
* tree-vect-stmts.c (vectorizable_load): Fix error in
previous change with regard to gather vectorization.

2 years agoarm/66791: Replace builtins for vdup_n and vmov_n intrinsics.
prathamesh.kulkarni [Wed, 11 Aug 2021 10:00:14 +0000 (15:30 +0530)]
arm/66791: Replace builtins for vdup_n and vmov_n intrinsics.

gcc/ChangeLog:

PR target/66791
* config/arm/arm_neon.h (vdup_n_s8): Replace call to builtin
with constructor.
(vdup_n_s16): Likewise.
(vdup_n_s32): Likewise.
(vdup_n_s64): Likewise.
(vdup_n_u8): Likewise.
(vdup_n_u16): Likewise.
(vdup_n_u32): Likewise.
(vdup_n_u64): Likewise.
(vdup_n_p8): Likewise.
(vdup_n_p16): Likewise.
(vdup_n_p64): Likewise.
(vdup_n_f16): Likewise.
(vdup_n_f32): Likewise.
(vdupq_n_s8): Likewise.
(vdupq_n_s16): Likewise.
(vdupq_n_s32): Likewise.
(vdupq_n_s64): Likewise.
(vdupq_n_u8): Likewise.
(vdupq_n_u16): Likewise.
(vdupq_n_u32): Likewise.
(vdupq_n_u64): Likewise.
(vdupq_n_p8): Likewise.
(vdupq_n_p16): Likewise.
(vdupq_n_p64): Likewise.
(vdupq_n_f16): Likewise.
(vdupq_n_f32): Likewise.
(vmov_n_s8): Replace call to builtin with call to corresponding
vdup_n intrinsic.
(vmov_n_s16): Likewise.
(vmov_n_s32): Likewise.
(vmov_n_s64): Likewise.
(vmov_n_u8): Likewise.
(vmov_n_u16): Likewise.
(vmov_n_u32): Likewise.
(vmov_n_u64): Likewise.
(vmov_n_p8): Likewise.
(vmov_n_p16): Likewise.
(vmov_n_f16): Likewise.
(vmov_n_f32): Likewise.
(vmovq_n_s8): Likewise.
(vmovq_n_s16): Likewise.
(vmovq_n_s32): Likewise.
(vmovq_n_s64): Likewise.
(vmovq_n_u8): Likewise.
(vmovq_n_u16): Likewise.
(vmovq_n_u32): Likewise.
(vmovq_n_u64): Likewise.
(vmovq_n_p8): Likewise.
(vmovq_n_p16): Likewise.
(vmovq_n_f16): Likewise.
(vmovq_n_f32): Likewise.
* config/arm/arm_neon_builtins.def: Remove entries for vdup_n.

gcc/testsuite/ChangeLog:

PR target/66791
* gcc.target/arm/pr51534.c: Adjust test.

2 years agosanitizer: Cherry-pick realpath fix
Jakub Jelinek [Wed, 11 Aug 2021 08:23:34 +0000 (10:23 +0200)]
sanitizer: Cherry-pick realpath fix

tsan in some cases starts ignoring interceptors and only calls the
intercepted functions.  But for realpath the behavior for NULL second argument
was only handled in the interceptor and intercepted function was the one
found by dlsym which is often one that doesn't handle NULL as second argument.

Fixed by using dlvsym with "GLIBC_2.3" if possible for intercepted function
and don't emulate behavior in the wrapper.

2021-08-11  Jakub Jelinek  <jakub@redhat.com>

* sanitizer_common/sanitizer_common_interceptors.inc: Cherry-pick
llvm-project revision faef0d042f523357fe5590e7cb6a8391cf0351a8.

2 years agoAda: Remove debug line number for DECL_IGNORED_P functions
Bernd Edlinger [Sat, 24 Jul 2021 10:53:39 +0000 (12:53 +0200)]
Ada: Remove debug line number for DECL_IGNORED_P functions

It was pointed out in PR101598 to be inappropriate, that
ignored Ada decls receive the source line number which was
recorded in the function decl's DECL_SOURCE_LOCATION.
Therefore set all front-end-generated Ada decls with
DECL_IGNORED_P to UNKNOWN_LOCATION.

2021-08-11  Bernd Edlinger  <bernd.edlinger@hotmail.de>

PR debug/101598
* gcc-interface/trans.c (Subprogram_Body_to_gnu): Set the
DECL_SOURCE_LOCATION of DECL_IGNORED_P gnu_subprog_decl to
UNKNOWN_LOCATION.

2 years agocompiler: don't crash on a, b := int(0)
Ian Lance Taylor [Tue, 10 Aug 2021 23:13:01 +0000 (16:13 -0700)]
compiler: don't crash on a, b := int(0)

Fixes PR go/101851

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/341330

2 years agoExtend ldexp{s,d}f3 to vscalefs{s,d} when TARGET_AVX512F and TARGET_SSE_MATH.
liuhongt [Tue, 10 Aug 2021 11:00:18 +0000 (19:00 +0800)]
Extend ldexp{s,d}f3 to vscalefs{s,d} when TARGET_AVX512F and TARGET_SSE_MATH.

gcc/ChangeLog:

PR target/98309
* config/i386/i386.md (ldexp<mode>3): Extend to vscalefs[sd]
when TARGET_AVX512F and TARGET_SSE_MATH.

gcc/testsuite/ChangeLog:

PR target/98309
* gcc.target/i386/pr98309-1.c: New test.
* gcc.target/i386/pr98309-2.c: New test.

2 years agogcc.dg/uninit-pred-9_b.c: Xfail for CRIS too
Hans-Peter Nilsson [Tue, 10 Aug 2021 23:40:12 +0000 (01:40 +0200)]
gcc.dg/uninit-pred-9_b.c: Xfail for CRIS too

Adding to the growing list, for autotester accounting purposes.

FWIW I see this fails for m68k too:
https://gcc.gnu.org/pipermail/gcc-testresults/2021-August/712395.html
and moxie:
https://gcc.gnu.org/pipermail/gcc-testresults/2021-August/712389.html
and pru:
https://gcc.gnu.org/pipermail/gcc-testresults/2021-August/712366.html

testsuite:
PR middle-end/101674
* gcc.dg/uninit-pred-9_b.c: Xfail for cris-*-* too.

2 years agoDaily bump.
GCC Administrator [Wed, 11 Aug 2021 00:16:27 +0000 (00:16 +0000)]
Daily bump.

2 years agoopenmp: Fix up cp/parser.c build with GCC 4.8 to 6
Jakub Jelinek [Tue, 10 Aug 2021 16:01:23 +0000 (18:01 +0200)]
openmp: Fix up cp/parser.c build with GCC 4.8 to 6

Christophe Lyon reported that cp/parser.c no longer compiles with
GCC 4.8.5 after my recent OpenMP changes.
A goto out; there crosses odsd variable declaration, and odsd has
a vec<...> member where vec has = default; default constructor
and gcc before r7-2822-gd0b0fbd9fce2f30a82558bf2308b3a7b56c2f364
treated that as error.

Fixed by moving the declaration earlier before the goto.

Tested on x86_64-linux with GCC 4.8.5 system gcc, committed to trunk
as obvious.

2021-08-10  Jakub Jelinek  <jakub@redhat.com>

* parser.c (cp_parser_member_declaration): Move odsd declaration
before cp_parser_using_declaration call to avoid errors with
GCC 4.8 to 6.

2 years agomklog: support '-b c/101343' format.
Martin Liska [Mon, 12 Jul 2021 10:07:19 +0000 (12:07 +0200)]
mklog: support '-b c/101343' format.

contrib/ChangeLog:

* mklog.py: Support additional PRs without PR prefix.

2 years agogfortran: Fix in-build-tree testing [PR101305, PR101660]
Tobias Burnus [Tue, 10 Aug 2021 15:26:32 +0000 (17:26 +0200)]
gfortran: Fix in-build-tree testing [PR101305, PR101660]

ISO_Fortran_binding.h is written in the build dir - hence, a previous commit
added it as include directory for in-build-tree testing.  However,
it turned out that -I$specdir/libgfortran interferes with reading .mod files
as they are then no longer regareded as intrinsic modules.  Solution: Create
an extra include/ directory in the libgfortran build dir and copy
ISO_Fortran_binding.h to that directory.  As -B$specdir/libgfortran already
causes gfortran to read that include subdirectory, the -I flag is no longer
needed.

PR libfortran/101305
PR fortran/101660
PR testsuite/101847

libgfortran/ChangeLog:

* Makefile.am (ISO_Fortran_binding.h): Create include/ in the build dir
and copy the include file to it.
(clean-local): Add for removing the 'include' directory.
* Makefile.in: Regenerate.

gcc/testsuite/ChangeLog:

* lib/gfortran.exp (gfortran_init): Remove -I$specpath/libgfortran
from the string used to set GFORTRAN_UNDER_TEST.

2 years agoEnable gcc.target/i386/pr88531-1a.c for all targets
H.J. Lu [Tue, 10 Aug 2021 12:30:44 +0000 (05:30 -0700)]
Enable gcc.target/i386/pr88531-1a.c for all targets

PR tree-optimization/101809
* gcc.target/i386/pr88531-1a.c: Enable for all targets.

2 years agoi386: Allow some V32HImode and V64QImode permutations even without AVX512BW [PR80355]
Jakub Jelinek [Tue, 10 Aug 2021 10:38:00 +0000 (12:38 +0200)]
i386: Allow some V32HImode and V64QImode permutations even without AVX512BW [PR80355]

When working on the PR, I've noticed we generate terrible code for
V32HImode or V64QImode permutations for -mavx512f -mno-avx512bw.
Generally we can't do much with such permutations, but since PR68655
we can handle at least some, those expressible using V16SImode or V8DImode
permutations, but that wasn't reachable, because ix86_vectorize_vec_perm_const
didn't even try, it said without TARGET_AVX512BW it can't do anything, and
with it can do everything, no d.testing_p attempts.

This patch makes it try it for TARGET_AVX512F && !TARGET_AVX512BW.

The first hunk is to avoid ICE, expand_vec_perm_even_odd_1 asserts d->vmode
isn't V32HImode because expand_vec_perm_1 for AVX512BW handles already
all permutations, but when we let it through without !TARGET_AVX512BW,
expand_vec_perm_1 doesn't handle it.

If we want, that hunk can be dropped if we implement in
expand_vec_perm_even_odd_1 and its helper the even permutation as
vpmovdw + vpmovdw + vinserti64x4 and odd permutation as
vpsrld $16 + vpsrld $16 + vpmovdw + vpmovdw + vinserti64x4.

2021-08-10  Jakub Jelinek  <jakub@redhat.com>

PR target/80355
* config/i386/i386-expand.c (expand_vec_perm_even_odd): Return false
for V32HImode if !TARGET_AVX512BW.
(ix86_vectorize_vec_perm_const) <case E_V32HImode, case E_V64QImode>:
If !TARGET_AVX512BW and TARGET_AVX512F and d.testing_p, don't fail
early, but actually check the permutation.

* gcc.target/i386/avx512f-pr80355-2.c: New test.

2 years agotree-optimization/101809 - support emulated gather for double[int]
Richard Biener [Tue, 10 Aug 2021 08:54:58 +0000 (10:54 +0200)]
tree-optimization/101809 - support emulated gather for double[int]

This adds emulated gather support for index vectors with more
elements than the data vector.  The internal function gather
vectorization code doesn't currently handle this (but the builtin
decl code does).  This allows vectorization of double data gather
with int indexes on 32bit platforms where there isn't an implicit
widening to 64bit present.

2021-08-10  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101809
* tree-vect-stmts.c (get_load_store_type): Allow emulated
gathers with offset vector nunits being a constant multiple
of the data vector nunits.
(vect_get_gather_scatter_ops): Use the appropriate nunits
for the offset vector defs.
(vectorizable_store): Adjust call to
vect_get_gather_scatter_ops.
(vectorizable_load): Likewise.  Handle the case of less
offset vectors than data vectors.

2 years agoi386: Improve single operand AVX512F permutations [PR80355]
Jakub Jelinek [Tue, 10 Aug 2021 09:34:53 +0000 (11:34 +0200)]
i386: Improve single operand AVX512F permutations [PR80355]

On the following testcase we emit
vmovdqa32 .LC0(%rip), %zmm1
vpermd %zmm0, %zmm1, %zmm0
and
vmovdqa64 .LC1(%rip), %zmm1
vpermq %zmm0, %zmm1, %zmm0
instead of
vshufi32x4 $78, %zmm0, %zmm0, %zmm0
and
vshufi64x2 $78, %zmm0, %zmm0, %zmm0
we can emit with the patch.  We have patterns that match two argument
permutations for vshuf[if]*, but for one argument it doesn't trigger.
Either we can add two patterns for that, or we would need to add another
routine to i386-expand.c that would transform under certain condition
these cases to the two argument vshuf*, doing it in sse.md looked simpler.
We don't need this for 32-byte vectors, we already emit single insn
permutation that doesn't need memory op there.

2021-08-10  Jakub Jelinek  <jakub@redhat.com>

PR target/80355
* config/i386/sse.md (*avx512f_shuf_<shuffletype>64x2_1<mask_name>_1,
*avx512f_shuf_<shuffletype>32x4_1<mask_name>_1): New define_insn
patterns.

* gcc.target/i386/avx512f-pr80355-1.c: New test.

2 years agoopenmp: Add support for declare simd and declare variant in a attribute syntax
Jakub Jelinek [Tue, 10 Aug 2021 09:22:33 +0000 (11:22 +0200)]
openmp: Add support for declare simd and declare variant in a attribute syntax

This patch adds support for declare simd and declare variant in attribute
syntax.  Either in attribute-specifier-seq at the start of declaration, in
that case it has similar restriction to pragma-syntax, that there is a single
function declaration/definition in the declaration, rather than variable
declaration or more than one function declarations or mix of function and
variable declarations.  Or after the declarator id, in that case it applies
just to the single function declaration and the same declaration can have
multiple such attributes.  Or both.

Furthermore, cp_parser_statement has been adjusted so that it doesn't
accept [[omp::directive (parallel)]] etc. before statements that don't
take attributes at all, or where those attributes don't appertain to
the statement but something else (e.g. to label, using directive,
declaration, etc.).

2021-08-10  Jakub Jelinek  <jakub@redhat.com>

gcc/cp/
* parser.h (struct cp_omp_declare_simd_data): Remove
in_omp_attribute_pragma and clauses members, add loc and attribs.
(struct cp_oacc_routine_data): Remove loc member, add clauses
member.
* parser.c (cp_finalize_omp_declare_simd): New function.
(cp_parser_handle_statement_omp_attributes): Mention in
function comment the function is used also for
attribute-declaration.
(cp_parser_handle_directive_omp_attributes): New function.
(cp_parser_statement): Don't call
cp_parser_handle_statement_omp_attributes if statement doesn't
have attribute-specifier-seq at the beginning at all or if
if those attributes don't appertain to the statement.
(cp_parser_simple_declaration): Call
cp_parser_handle_directive_omp_attributes and
cp_finalize_omp_declare_simd.
(cp_parser_explicit_instantiation): Likewise.
(cp_parser_init_declarator): Initialize prefix_attributes
only after parsing declarators.
(cp_parser_direct_declarator): Call
cp_parser_handle_directive_omp_attributes and
cp_finalize_omp_declare_simd.
(cp_parser_member_declaration): Likewise.
(cp_parser_single_declaration): Likewise.
(cp_parser_omp_declare_simd): Don't initialize
data.in_omp_attribute_pragma, instead initialize
data.attribs[0] and data.attribs[1].
(cp_finish_omp_declare_variant): Remove
in_omp_attribute_pragma argument, instead use
parser->lexer->in_omp_attribute_pragma.
(cp_parser_late_parsing_omp_declare_simd): Adjust
cp_finish_omp_declare_variant caller.  Handle attribute-syntax
declare simd/variant.
gcc/testsuite/
* g++.dg/gomp/attrs-1.C (bar): Add missing semicolon after
[[omp::directive (threadprivate (t2))]].  Add tests with
if/while/switch after parallel in attribute syntax.
(corge): Add missing omp:: before directive.
* g++.dg/gomp/attrs-2.C (bar): Add missing semicolon after
[[omp::directive (threadprivate (t2))]].
* g++.dg/gomp/attrs-10.C: New test.
* g++.dg/gomp/attrs-11.C: New test.

2 years agoi386: Fix typos in amxbf16 runtime test.
Hongyu Wang [Tue, 10 Aug 2021 07:13:25 +0000 (15:13 +0800)]
i386: Fix typos in amxbf16 runtime test.

gcc/testsuite/ChangeLog:

* gcc.target/i386/amxbf16-dpbf16ps-2.c: Fix typos.

2 years agotree-optimization/101801 - rework generic vector vectorization more
Richard Biener [Mon, 9 Aug 2021 09:42:47 +0000 (11:42 +0200)]
tree-optimization/101801 - rework generic vector vectorization more

This builds ontop of the vect_worthwhile_without_simd_p refactoring
done earlier.  It was wrong in dropping the appearant double checks
for operation support since the optab check can happen with an
integer vector emulation mode and thus succeed but vector lowering
might not actually support the operation on word_mode.

The following patch adds a vect_emulated_vector_p helper and
re-instantiates the check where it was previously.  It also adds
appropriate costing of the scalar stmts emitted by vector lowering
to vectorizable_operation which should be the only place such
operations are synthesized.  I've also cared for the case where
the vector mode is supported but the operation is not (though
I think this will be unlikely given we're talking about plus, minus
and negate).

This fixes the observed FAIL of gcc.dg/tree-ssa/gen-vect-11b.c
with -m32 where we end up vectorizing a multiplication that ends up
being teared down to scalars again by vector lowering.

I'm not super happy about all the other places where we're now
and previously feeding scalar modes to optab checks where we
want to know whether we can vectorize sth but well.

2021-09-08  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101801
PR tree-optimization/101819
* tree-vectorizer.h (vect_emulated_vector_p): Declare.
* tree-vect-loop.c (vect_emulated_vector_p): New function.
(vectorizable_reduction): Re-instantiate a check for emulated
operations.
* tree-vect-stmts.c (vectorizable_shift): Likewise.
(vectorizable_operation): Likewise.  Cost emulated vector
operations according to the scalar sequence synthesized by
vector lowering.

2 years agomiddle-end/101824 - properly handle volatiles in nested fn lowering
Richard Biener [Mon, 9 Aug 2021 08:19:10 +0000 (10:19 +0200)]
middle-end/101824 - properly handle volatiles in nested fn lowering

When we build the COMPONENT_REF of a formerly volatile local off
the FRAME decl we have to make sure to mark the COMPONENT_REF
as TREE_THIS_VOLATILE.  While the GIMPLE operand scanner looks
at the FIELD_DECL this is not how volatile GENERIC refs work.

2021-08-09  Richard Biener  <rguenther@suse.de>

PR middle-end/101824
* tree-nested.c (get_frame_field): Mark the COMPONENT_REF as
volatile in case the variable was.

* gcc.dg/tree-ssa/pr101824.c: New testcase.

2 years agoEvaluate arguments of sizeof that are structs of variable size.
Martin Uecker [Tue, 10 Aug 2021 05:42:51 +0000 (07:42 +0200)]
Evaluate arguments of sizeof that are structs of variable size.

Evaluate arguments of sizeof for all types of variable size
and not just for VLAs. This fixes some issues related to
[PR29970] where statement expressions need to be evaluated
so that the size is well defined.

2021-08-10  Martin Uecker  <muecker@gwdg.de>

gcc/c/
PR c/29970
* c-typeck.c (c_expr_sizeof_expr): Evaluate
size expressions for structs of variable size.

gcc/testsuite/
PR c/29970
* gcc.dg/vla-stexp-1.c: New test.

2 years agox86: Optimize load of const FP all bits set vectors
H.J. Lu [Fri, 6 Aug 2021 19:32:01 +0000 (12:32 -0700)]
x86: Optimize load of const FP all bits set vectors

Check float_vector_all_ones_operand for vector floating-point modes to
optimize load of const floating-point all bits set vectors.

gcc/

PR target/101804
* config/i386/constraints.md (BC): Document for integer SSE
constant all bits set operand.
(BF): New constraint for const floating-point all bits set
vectors.
* config/i386/i386.c (standard_sse_constant_p): Likewise.
(standard_sse_constant_opcode): Likewise.
* config/i386/sse.md (sseconstm1): New mode attribute.
(mov<mode>_internal): Replace BC with <sseconstm1>.

gcc/testsuite/

PR target/101804
* gcc.target/i386/avx2-gather-2.c: Pass -march=skylake instead
of "-mavx2 -mtune=skylake".  Scan vpcmpeqd.

2 years agoSupport cond_ashr/lshr/ashl for vector integer modes under AVX512.
liuhongt [Thu, 5 Aug 2021 09:51:48 +0000 (17:51 +0800)]
Support cond_ashr/lshr/ashl for vector integer modes under AVX512.

gcc/ChangeLog:

* config/i386/sse.md (cond_<insn><mode>): New expander.
(VI248_AVX512VLBW): New mode iterator.
* config/i386/predicates.md
(nonimmediate_or_const_vec_dup_operand): New predicate.

gcc/testsuite/ChangeLog:

* gcc.target/i386/cond_op_shift_d-1.c: New test.
* gcc.target/i386/cond_op_shift_d-2.c: New test.
* gcc.target/i386/cond_op_shift_q-1.c: New test.
* gcc.target/i386/cond_op_shift_q-2.c: New test.
* gcc.target/i386/cond_op_shift_ud-1.c: New test.
* gcc.target/i386/cond_op_shift_ud-2.c: New test.
* gcc.target/i386/cond_op_shift_uq-1.c: New test.
* gcc.target/i386/cond_op_shift_uq-2.c: New test.
* gcc.target/i386/cond_op_shift_uw-1.c: New test.
* gcc.target/i386/cond_op_shift_uw-2.c: New test.
* gcc.target/i386/cond_op_shift_w-1.c: New test.
* gcc.target/i386/cond_op_shift_w-2.c: New test.

2 years agoDaily bump.
GCC Administrator [Tue, 10 Aug 2021 00:16:28 +0000 (00:16 +0000)]
Daily bump.

2 years agoEnsure toupper and tolower follow the expected pattern.
Andrew MacLeod [Mon, 9 Aug 2021 19:53:42 +0000 (15:53 -0400)]
Ensure toupper and tolower follow the expected pattern.

If the parameter is not compatible with the LHS, assume this is not really a
builtin function to avoid a trap.

gcc/
PR tree-optimization/101741
* gimple-range-fold.cc (fold_using_range::range_of_builtin_call): Check
type of parameter for toupper/tolower.

gcc/testsuite/
* gcc.dg/pr101741.c: New.

2 years agolibstdc++: Reduce use of debug containers in <regex>
Jonathan Wakely [Mon, 9 Aug 2021 10:49:09 +0000 (11:49 +0100)]
libstdc++: Reduce use of debug containers in <regex>

The std::regex code uses std::map and std::vector, which means that when
_GLIBCXX_DEBUG is defined it uses the debug versions of those
containers. That no longer compiles, because I changed <regex> to
include <bits/stl_map.h> and <bits/stl_vector.h> instead of <map> and
<vector>, so the debug versions aren't defined, and std::map doesn't
compile. There is also a use of std::stack, which defaults to std::deque
which is the debug deque when _GLIBCXX_DEBUG is defined.

Using std::map, std::vector, and std::deque is probably a mistake, and
we should qualify them with _GLIBCXX_STD_C instead so that the debug
versions aren't used. We do not need the overhead of checking our own
uses of those containers, which should be correct anyway. The exception
is the vector base class of std::match_results, which exposes iterators
to users, so can benefit from debug mode checks for its iterators. For
other accesses to the vector elements, match_results already does its
own checks, so can access the _GLIBCXX_STD_C::vector base class
directly.

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

* include/bits/regex.h (basic_regex::transform_primary): Use
_GLIBCXX_STD_C::vector for local variable.
* include/bits/regex.tcc (__regex_algo_impl): Use reference to
_GLIBCXX_STD_C::vector base class of match_results.
* include/bits/regex_automaton.tcc (_StateSeq:_M_clone): Use
_GLIBCXX_STD_C::map and _GLIBCXX_STD_C::deque for local
variables.
* include/bits/regex_compiler.h (_BracketMatcher): Use
_GLIBCXX_STD_C::vector for data members.
* include/bits/regex_executor.h (_Executor): Likewise.
* include/std/regex [_GLIBCXX_DEBUG]: Include <debug/vector>.

2 years agolibstdc++: [_GLIBCXX_DEBUG] Avoid allocator operator== when always equal
François Dumont [Mon, 9 Aug 2021 09:55:43 +0000 (11:55 +0200)]
libstdc++: [_GLIBCXX_DEBUG] Avoid allocator operator== when always equal

Use std::allocator_traits::is_always_equal to find out if we need to compare
allocator instances on safe container allocator aware move constructor.

libstdc++-v3/ChangeLog:

* include/debug/safe_container.h
(_Safe_container(_Safe_container&&, const _Alloc&, std::true_type)): New.
(_Safe_container(_Safe_container&&, const _Alloc&, std::false_type)): New.
(_Safe_container(_Safe_container&&, const _Alloc&)): Use latters.

2 years agoipa: Fix testsuite/gcc.dg/ipa/remref-6.c
Martin Jambor [Mon, 9 Aug 2021 15:35:39 +0000 (17:35 +0200)]
ipa: Fix testsuite/gcc.dg/ipa/remref-6.c

I forgot to add -fdump-ipa-inline to options of
testsuite/gcc.dg/ipa/remref-6.c and so the dump scan test were not
PASSing but ended up as UNRESOLVED.  Fixing that revealed that the one
of the dumps it was looking for had a double space, so I removed it
too.

gcc/ChangeLog:

2021-08-09  Martin Jambor  <mjambor@suse.cz>

PR testsuite/101654
* ipa-prop.c (propagate_controlled_uses): Removed a spurious space.

gcc/testsuite/ChangeLog:

2021-08-09  Martin Jambor  <mjambor@suse.cz>

PR testsuite/101654
* gcc.dg/ipa/remref-6.c: Added missing -fdump-ipa-inline option.

2 years agoVerify destination[source] of a load[store] instruction is a register.
Pat Haugen [Mon, 9 Aug 2021 15:05:49 +0000 (10:05 -0500)]
Verify destination[source] of a load[store] instruction is a register.

gcc/ChangeLog:

* config/rs6000/rs6000.c (is_load_insn1): Verify destination is a
register.
(is_store_insn1): Verify source is a register.

2 years agoi386: Name V2SF logic insns [PR101812]
Uros Bizjak [Mon, 9 Aug 2021 14:38:54 +0000 (16:38 +0200)]
i386: Name V2SF logic insns [PR101812]

Name V2SF logic insns, so expand_simple_binop works with V2SF modes.

2021-08-09  Uroš Bizjak  <ubizjak@gmail.com>

gcc/
PR target/101812
* config/i386/mmx.md (<any_logic:code>v2sf3):
Rename from *mmx_<any_logic:code>v2sf3

gcc/testsuite/
PR target/101812
* gcc.target/i386/pr101812.c: New test.

2 years agoCross-reference parts adapted in 'gcc/omp-oacc-neuter-broadcast.cc'
Thomas Schwinge [Mon, 9 Aug 2021 10:21:43 +0000 (12:21 +0200)]
Cross-reference parts adapted in 'gcc/omp-oacc-neuter-broadcast.cc'

gcc/
* config/nvptx/nvptx.c: Cross-reference parts adapted in
'gcc/omp-oacc-neuter-broadcast.cc'.
* omp-low.c: Likewise.
* omp-oacc-neuter-broadcast.cc: Cross-reference parts adapted from
the above files.

2 years agoamdgcn: Enable OpenACC worker partitioning for AMD GCN
Julian Brown [Tue, 2 Mar 2021 12:20:13 +0000 (04:20 -0800)]
amdgcn: Enable OpenACC worker partitioning for AMD GCN

gcc/
* config/gcn/gcn.c (gcn_init_builtins): Override decls for
BUILT_IN_GOACC_SINGLE_START, BUILT_IN_GOACC_SINGLE_COPY_START,
BUILT_IN_GOACC_SINGLE_COPY_END and BUILT_IN_GOACC_BARRIER.
(gcn_goacc_validate_dims): Turn on worker partitioning unconditionally.
(gcn_fork_join): Update comment.
* config/gcn/gcn.opt (flag_worker_partitioning): Remove.
(macc_experimental_workers): Remove unused option.
libgomp/
* plugin/plugin-gcn.c (gcn_exec): Change default number of workers to
16.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-kernels-1.c
[acc_device_radeon]: Update.
* testsuite/libgomp.oacc-c-c++-common/loop-dim-default.c
[ACC_DEVICE_TYPE_radeon]: Likewise.
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c
[acc_device_radeon]: Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-wv-2.c
[ACC_DEVICE_TYPE_radeon]: Likewise.
* testsuite/libgomp.oacc-fortran/optional-reduction.f90: XFAIL for
'openacc_radeon_accel_selected' and '-O0'.
* testsuite/libgomp.oacc-fortran/reduction-7.f90: Likewise.

Co-Authored-By: Kwok Cheung Yeung <kcy@codesourcery.com>
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
2 years agoopenacc: Middle-end worker-partitioning support
Julian Brown [Tue, 2 Mar 2021 12:20:11 +0000 (04:20 -0800)]
openacc: Middle-end worker-partitioning support

This patch implements worker-partitioning support in the middle end,
by rewriting gimple. The OpenACC execution model requires that code
can run in either "worker single" mode where only a single worker per
gang is active, or "worker partitioned" mode, where multiple workers
per gang are active. This means we need to do something equivalent
to spawning additional workers when transitioning from worker-single
to worker-partitioned mode. However, GPUs typically fix the number of
threads of invoked kernels at launch time, so we need to do something
with the "extra" threads when they are not wanted.

The scheme used is to conditionalise each basic block that executes
in "worker single" mode for worker 0 only. Conditional branches
are handled specially so "idle" (non-0) workers follow along with
worker 0. On transitioning to "worker partitioned" mode, any variables
modified by worker 0 are propagated to the other workers via GPU shared
memory. Special care is taken for routine calls, writes through pointers,
and so forth, as follows:

  - There are two types of function calls to consider in worker-single
    mode: "normal" calls to maths library routines, etc. are called from
    worker 0 only. OpenACC routines may contain worker-partitioned loops
    themselves, so are called from all workers, including "idle" ones.

  - SSA names set in worker-single mode, but used in worker-partitioned
    mode, are copied to shared memory in worker 0. Other workers retrieve
    the value from the appropriate shared-memory location after a barrier,
    and new phi nodes are introduced at the convergence point to resolve
    the worker 0/other worker copies of the value.

  - Local scalar variables (on the stack) also need special handling. We
    broadcast any variables that are written in the current worker-single
    block, and that are read in any worker-partitioned block.  (This is
    believed to be safe, and is flow-insensitive to ease analysis.)

  - Local aggregates (arrays and composites) on the stack are *not*
    broadcast. Instead we force gimple stmts modifying elements/fields of
    local aggregates into fully-partitioned mode. The RHS of the
    assignment is a scalar, and is thus subject to broadcasting as above.

  - Writes through pointers may affect any local variable that has
    its address taken. We use points-to analysis to determine the set
    of potentially-affected variables for a given pointer indirection.
    We broadcast any such variable which is used in worker-partitioned
    mode, on a per-block basis for any block containing a write through
    a pointer.

Some slides about the implementation (from 2018) are available at:

  https://jtb20.github.io/gcnworkers.pdf

gcc/
* Makefile.in (OBJS): Add omp-oacc-neuter-broadcast.o.
* doc/tm.texi.in (TARGET_GOACC_CREATE_WORKER_BROADCAST_RECORD):
Add documentation hook.
* doc/tm.texi: Regenerate.
* omp-oacc-neuter-broadcast.cc: New file.
* omp-builtins.def (BUILT_IN_GOACC_BARRIER)
(BUILT_IN_GOACC_SINGLE_START, BUILT_IN_GOACC_SINGLE_COPY_START)
(BUILT_IN_GOACC_SINGLE_COPY_END): New builtins.
* passes.def (pass_omp_oacc_neuter_broadcast): Add pass.
* target.def (goacc.create_worker_broadcast_record): Add target
hook.
* tree-pass.h (make_pass_omp_oacc_neuter_broadcast): Add
prototype.
* config/gcn/gcn-protos.h (gcn_goacc_adjust_propagation_record):
Rename prototype to...
(gcn_goacc_create_worker_broadcast_record): ... this.
* config/gcn/gcn-tree.c (gcn_goacc_adjust_propagation_record): Rename
function to...
(gcn_goacc_create_worker_broadcast_record): ... this.
* config/gcn/gcn.c (TARGET_GOACC_ADJUST_PROPAGATION_RECORD):
Rename to...
(TARGET_GOACC_CREATE_WORKER_BROADCAST_RECORD): ... this.

Co-Authored-By: Nathan Sidwell <nathan@codesourcery.com> (via 'gcc/config/nvptx/nvptx.c' master)
Co-Authored-By: Kwok Cheung Yeung <kcy@codesourcery.com>
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
2 years agoPR101609: Use the correct iterator for AArch64 vector right shift pattern
Tejas Belagod [Mon, 9 Aug 2021 10:33:30 +0000 (11:33 +0100)]
PR101609: Use the correct iterator for AArch64 vector right shift pattern

Loops containing long long shifts fail to vectorize due to the vectorizer
not being able to recognize long long right shifts. This is due to a bug
in the iterator used for the vashr and vlshr patterns in aarch64-simd.md.

2021-08-09  Tejas Belagod  <tejas.belagod@arm.com>

gcc/ChangeLog
PR target/101609
* config/aarch64/aarch64-simd.md (vlshr<mode>3, vashr<mode>3): Use
the right iterator.

gcc/testsuite/ChangeLog
* gcc.target/aarch64/vect-shr-reg.c: New testcase.
* gcc.target/aarch64/vect-shr-reg-run.c: Likewise.

2 years agoRemove 'gcc/omp-offload.c' from 'GTFILES'
Thomas Schwinge [Wed, 4 Aug 2021 11:41:22 +0000 (13:41 +0200)]
Remove 'gcc/omp-offload.c' from 'GTFILES'

Given that it doesn't contain any 'GTY' markers, no 'gcc/gt-omp-offload.h' file
gets generated (and '#include'd anywhere).

Small fix-up for r243673 (Git commit 629b3d75c8c5a244d891a9c292bca6912d4b0dd9)
"Split omp-low into multiple files".

gcc/
* Makefile.in (GTFILES): Remove '$(srcdir)/omp-offload.c'.

2 years agoDon't consider '-foffload-abi' in 'DEF_GOACC_BUILTIN', 'DEF_GOMP_BUILTIN'
Thomas Schwinge [Mon, 2 Aug 2021 16:33:50 +0000 (18:33 +0200)]
Don't consider '-foffload-abi' in 'DEF_GOACC_BUILTIN', 'DEF_GOMP_BUILTIN'

Since Tom's PR64707 commit r220037 (Git commit
1506ae0e1e865fb7a42fc37a47f1799b71f21c53) "Make fopenmp an LTO option" as well
as PR64672 commit r220038 (Git commit a0c88d0629a33161add8d5bc083f1e59f3f756f7)
"Make fopenacc an LTO option", we're now actually passing
'-fopenacc'/'-fopenmp' to the 'mkoffload's, which will pass these on to the
offload compilers.

gcc/
* builtins.def (DEF_GOACC_BUILTIN, DEF_GOMP_BUILTIN): Don't
consider '-foffload-abi'.
* common.opt (-foffload-abi): Remove 'Var', 'Init'.
* opts.c (common_handle_option) <-foffload-abi> [ACCEL_COMPILER]:
Ignore.