Add wi_fold_in_parts.

range-ops uses wi_fold to individually fold subranges one at a time and
then combined them.  This patch first calls wi_fold_in_parts which checks if
one of the subranges is small, and if so, further splits that subrange
into constants.

gcc/
PR tree-optimization/96542
* range-op.cc (range_operator::wi_fold_in_parts): New.
(range_operator::fold_range): Call wi_fold_in_parts.
(operator_lshift::wi_fold): Fix broken lshift by [0,0].
* range-op.h (wi_fold_in_parts): Add prototype.

gcc/testsuite
* gcc.dg/pr96542.c: New.

Daily bump.

libstdc++: Improve diagnostics for std::get with invalid tuple index

This adds a deleted overload of std::get<I>(const tuple<Types...>&).
Invalid calls with an out of range index will match the deleted overload
and give a single, clear error about calling a deleted function, instead
of overload resolution errors for every std::get overload in the
library.

This changes the current output of 15+ errors (plus notes and associated
header context) into just two errors (plus context):

error: static assertion failed: tuple index must be in range
error: use of deleted function 'constexpr std::__enable_if_t<(__i >= sizeof... (_Types))> std::get(const std::tuple<_Types ...>&) [with long unsigned int __i = 1; _Elements = {int}; std::__enable_if_t<(__i >= sizeof... (_Types))> = void]'

This seems like a nice improvement, although PR c++/66968 means that
"_Types" is printed in the signature rather than "_Elements".

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

* include/std/tuple (get<I>): Add deleted overload for bad
index.
* testsuite/20_util/tuple/element_access/get_neg.cc: Adjust
expected errors.

c++: alias CTAD in unevaluated context [PR101233]

This is the alias CTAD version of the CTAD bug PR93248, and the fix is
the same: clear cp_unevaluated_operand so that the entire chain of
DECL_ARGUMENTS gets substituted.

PR c++/101233

gcc/cp/ChangeLog:

* pt.c (alias_ctad_tweaks): Clear cp_unevaluated_operand for
substituting DECL_ARGUMENTS.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/class-deduction-alias10.C: New test.

c++: covariant reference return types [PR99664]

This implements the wording changes of CWG 960 which clarifies that two
reference types are covariant only if they're both lvalue references
or both rvalue references.

DR 960
PR c++/99664

gcc/cp/ChangeLog:

* search.c (check_final_overrider): Compare TYPE_REF_IS_RVALUE
when the return types are references.

gcc/testsuite/ChangeLog:

* g++.dg/inherit/covariant23.C: New test.

analyzer: add region_model::check_region_access

I've been experimenting with various new diagnostics that
require a common place for the analyzer to check the validity
of reads or writes to memory (e.g. buffer overflow).

As preliminary work, this patch adds new
  region_model::check_region_for_{read|write} functions
which are called anywhere that the analyzer "sees" memory being
read from or written to (via region_model::get_store_value and
region_model::set_value).

This takes over the hardcoded calls to check_for_writable_region
(allowing for other kinds of checks on writes); checking reads is
currently a no-op.

gcc/analyzer/ChangeLog:
* analyzer.h (enum access_direction): New.
* engine.cc (exploded_node::on_longjmp): Update for new param of
get_store_value.
* program-state.cc (program_state::prune_for_point): Likewise.
* region-model-impl-calls.cc (region_model::impl_call_memcpy):
Replace call to check_for_writable_region with call to
check_region_for_write.
(region_model::impl_call_memset): Likewise.
(region_model::impl_call_strcpy): Likewise.
* region-model-reachability.cc (reachable_regions::add): Update
for new param of get_store_value.
* region-model.cc (region_model::get_rvalue_1): Likewise, also for
get_rvalue_for_bits.
(region_model::get_store_value): Add ctxt param and use it to call
check_region_for_read.
(region_model::get_rvalue_for_bits): Add ctxt param and use it to
call get_store_value.
(region_model::check_region_access): New.
(region_model::check_region_for_write): New.
(region_model::check_region_for_read): New.
(region_model::set_value): Update comment.  Replace call to
check_for_writable_region with call to check_region_for_write.
* region-model.h (region_model::get_rvalue_for_bits): Add ctxt
param.
(region_model::get_store_value): Add ctxt param.
(region_model::check_region_access): New decl.
(region_model::check_region_for_write): New decl.
(region_model::check_region_for_read): New decl.
* region.cc (region_model::copy_region): Update call to
get_store_value.
* svalue.cc (initial_svalue::implicitly_live_p): Likewise.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

analyzer: add __analyzer_dump_state

gcc/analyzer/ChangeLog:
* engine.cc (exploded_node::on_stmt_pre): Handle
__analyzer_dump_state.
* program-state.cc (extrinsic_state::get_sm_idx_by_name): New.
(program_state::impl_call_analyzer_dump_state): New.
* program-state.h (extrinsic_state::get_sm_idx_by_name): New decl.
(program_state::impl_call_analyzer_dump_state): New decl.
* region-model-impl-calls.cc
(call_details::get_arg_string_literal): New.
* region-model.h (call_details::get_arg_string_literal): New decl.

gcc/ChangeLog:
* doc/analyzer.texi: Add __analyzer_dump_state.

gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/analyzer-decls.h (__analyzer_dump_state): New.
* gcc.dg/analyzer/dump-state.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

analyzer: add svalue::maybe_get_region

gcc/analyzer/ChangeLog:
* program-state.cc (program_state::detect_leaks): Simplify using
svalue::maybe_get_region.
* region-model-impl-calls.cc (region_model::impl_call_fgets): Likewise.
(region_model::impl_call_fread): Likewise.
(region_model::impl_call_free): Likewise.
(region_model::impl_call_operator_delete): Likewise.
* region-model.cc (selftest::test_stack_frames): Likewise.
(selftest::test_state_merging): Likewise.
* svalue.cc (svalue::maybe_get_region): New.
* svalue.h (svalue::maybe_get_region): New decl.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

Fix PR 101453: ICE with optimize and large integer constant

The problem is the buffer is too small to hold "-O" and
the interger.  This fixes the problem by use the correct size
instead.

Changes since v1:
* v2: Use HOST_BITS_PER_LONG and just divide by 3 instead of
3.32.

OK? Bootstrapped and tested on x86_64-linux with no regressions.

gcc/c-family/ChangeLog:

PR c/101453
* c-common.c (parse_optimize_options): Use the correct
size for buffer.

PR testsuite/101468  - Wstringop-overflow tests failures

gcc/testsuite/ChangeLog:

PR testsuite/101468
* gcc.dg/Wstringop-overflow-41.c: Adjust to avoid target-specific
failures.
* gcc.dg/Wstringop-overflow-42.c: Same.
* gcc.dg/Wstringop-overflow-68.c: Same.
* gcc.dg/Wstringop-overflow-70.c: Same.
* gcc.dg/Wstringop-overflow-71.c: Same.
* gcc.dg/strlenopt-95.c: Fix typos.

rs6000: Red-black tree implementation for balanced tree search

2021-06-08  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rbtree.c: New file.
* config/rs6000/rbtree.h: New file.

rs6000: Add functions for matching types, part 3 of 3

2021-06-07  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-gen-builtins.c (restriction): New enum.
(typeinfo): Add restr field.
(match_bracketed_pair): New function.
(match_const_restriction): Implement.

rs6000: Add functions for matching types, part 2 of 3

2021-04-02  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-gen-builtins.c (match_basetype): Implement.

rs6000: Add functions for matching types, part 1 of 3

2021-06-07  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-gen-builtins.c (void_status): New enum.
(basetype): Likewise.
(typeinfo): Likewise.
(handle_pointer): New function.
(match_basetype): New stub function.
(match_const_restriction): Likewise.
(match_type): New function.

rs6000: Add helper functions for parsing

2021-07-16  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-gen-builtins.c (consume_whitespace): New
function.
(advance_line): Likewise.
(safe_inc_pos): Likewise.
(match_identifier): Likewise.
(match_integer): Likewise.
(match_to_right_bracket): Likewise.

rs6000: Add file support and functions for diagnostic support

2021-06-07  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-gen-builtins.c (bif_file): New variable.
(ovld_file): Likewise.
(header_file): Likewise.
(init_file): Likewise.
(defines_file): Likewise.
(pgm_path): Likewise.
(bif_path): Likewise.
(ovld_path): Likewise.
(header_path): Likewise.
(init_path): Likewise.
(defines_path): Likewise.
(LINELEN): New macro.
(linebuf): New variable.
(line): Likewise.
(pos): Likewise.
(diag): Likewise.
(bif_diag): New function.
(ovld_diag): Likewise.

rs6000: Add initial input files

This patch adds a tiny subset of the built-in and overload descriptions.

2021-04-02  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: New.
* config/rs6000/rs6000-overload.def: New.

rs6000: Initial create of rs6000-gen-builtins.c

2021-04-02  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-gen-builtins.c: New.

Support scanning of build-time GC roots in gengtype

Currently gengtype supports scanning target-specific files for GC roots,
but those files must exist in the source tree.  This patch extends the
support to include header files generated into the build directory.  It
also allows targets to specify build dependencies for s-gtype to ensure
the built headers are up to date prior to running gengtype.

2021-06-15  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* Makefile.in (EXTRA_GTYPE_DEPS): New variable.
(s-gtype): Depend on EXTRA_GTYPE_DEPS.
* gengtype-state.c (state_writer::write_state_file_list): Add a
parameter to the fileslist expression for the number of build
headers to scan.
(read_state_files_list): Detect build headers and strip the
initial "./" or ".\" from their names.
* gengtype.c (build_headers): New global variable.
(num_build_headers): Likewise.
(open_base_files): Emit #include for each build header.
(main): Detect and count build headers.
* gengtype.h (build_headers): New extern variable.
(num_build_headers): Likewise.

testsuite: fix IL32 issues with usdot tests.

Fix tests when int == long by using long long instead.

gcc/testsuite/ChangeLog:

PR middle-end/101457
* gcc.dg/vect/vect-reduc-dot-19.c: Use long long.
* gcc.dg/vect/vect-reduc-dot-20.c: Likewise.
* gcc.dg/vect/vect-reduc-dot-21.c: Likewise.
* gcc.dg/vect/vect-reduc-dot-22.c: Likewise.

libstdc++: Fix ChangeLog entry for commit r12-2326

libstdc++: Simplify numeric_limits<__max_size_type>

If __int128 is supported then __int_traits<__int128> is guaranteed to be
specialized, so we can remove the preprocessor condition inside the
std::numeric_traits<__detail::__max_size_type> specialization. Simply
using __int_traits<_Sp::__rep> gives the right answer.

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

* include/bits/max_size_type.h (numeric_limits<__max_size_type>):
Use __int_traits unconditionally.

libstdc++: Modernize <bits/random.h> helpers

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

* include/bits/random.h (_Shift::__value): Use constexpr.
(_Select_uint_least_t::type): Use using-declaration.
(_Mod): Likewise.
* testsuite/26_numerics/random/pr60037-neg.cc: Adjust dg-error
line number.

libstdc++: Use __extension__ instead of diagnostic pragmas

This reverts c1676651b6c417e8f2b276a28199d76943834277 and uses the
__extension__ keyword to prevent pedantic warnings instead of diagnostic
pragmas.

This also adds the __extension__ keyword in <limits> and <bits/random.h>
where there are some more warnings that I missed in the previous commit.

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

* include/bits/cpp_type_traits.h (__INT_N): Use __extension__
instead of diagnostic pragmas.
* include/bits/functional_hash.h: Likewise.
* include/bits/iterator_concepts.h (__is_signed_int128)
(__is_unsigned_int128): Likewise.
* include/bits/max_size_type.h (__max_size_type): Likewise.
(numeric_limits<__max_size_type>): Likewise.
* include/bits/std_abs.h (abs): Likewise.
* include/bits/stl_algobase.h (__size_to_integer): Likewise.
* include/bits/uniform_int_dist.h (uniform_int_distribution):
Likewise.
* include/ext/numeric_traits.h (_GLIBCXX_INT_N_TRAITS):
Likewise.
* include/std/type_traits (__is_integral_helper<INT_N>)
(__is_signed_integer, __is_unsigned_integer)
(__make_unsigned<INT_N>, __make_signed<INT_N>): Likewise.
* include/std/limits (__INT_N): Add __extension__ keyword.
* include/bits/random.h (_Select_uint_least_t)
(random_device): Likewise.

libstdc++: invalid default init in _CachedPosition [PR101231]

The primary template for _CachedPosition is a dummy implementation for
non-forward ranges, the iterators for which generally can't be cached.
Because this implementation doesn't actually cache anything, _M_has_value
is defined to be false and so calls to _M_get (which are always guarded
by _M_has_value) are unreachable.

Still, to suppress a "control reaches end of non-void function" warning
I made _M_get return {}, but after P2325 input iterators are no longer
necessarily default constructible so this workaround now breaks valid
programs.

This patch fixes this by instead using __builtin_unreachable to squelch
the warning.

PR libstdc++/101231

libstdc++-v3/ChangeLog:

* include/std/ranges (_CachedPosition::_M_get): For non-forward
ranges, just call __builtin_unreachable.
* testsuite/std/ranges/istream_view.cc (test05): New test.

libstdc++: Give split_view::_Sentinel a default ctor [PR101214]

This gives the new split_view's sentinel type a defaulted default
constructor, something which was overlooked in r12-1665.  This patch
also fixes a couple of other issues with the new split_view as reported
in the PR.

PR libstdc++/101214

libstdc++-v3/ChangeLog:

* include/std/ranges (split_view::split_view): Use std::move.
(split_view::_Iterator::_Iterator): Remove redundant
default_initializable constraint.
(split_view::_Sentinel::_Sentinel): Declare.
* testsuite/std/ranges/adaptors/split.cc (test02): New test.

c++: Don't hide narrowing errors in system headers

Jonathan pointed me at this issue where

  constexpr unsigned f() { constexpr int n = -1; return unsigned{n}; }

is accepted in system headers, despite the narrowing conversion from
a constant.  I suspect that whereas narrowing warnings should be
disabled, ill-formed narrowing of constants should be a hard error
(which can still be disabled by -Wno-narrowing).

gcc/cp/ChangeLog:

* typeck2.c (check_narrowing): Don't suppress the pedantic error
in system headers.

libstdc++-v3/ChangeLog:

* testsuite/20_util/ratio/operations/ops_overflow_neg.cc: Add
dg-error.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1y/Wnarrowing2.C: New test.
* g++.dg/cpp1y/Wnarrowing2.h: New test.

Remove more gimple_expr_type uses

This removes a few more uses.

2021-07-16  Richard Biener   <rguenther@suse.de>

* gimple-ssa-store-merging.c (verify_symbolic_number_p): Use
the type of the LHS.
(find_bswap_or_nop_1): Likewise.
(find_bswap_or_nop): Likewise.
* tree-vectorizer.h (vect_get_smallest_scalar_type): Adjust
prototype.
* tree-vect-data-refs.c (vect_get_smallest_scalar_type):
Remove unused parameters, pass in the scalar type.  Fix
internal store function handling.
* tree-vect-stmts.c (vect_analyze_stmt): Remove assert.
(vect_get_vector_types_for_stmt): Move down check for
existing vector stmt after we've determined a scalar type.
Pass down the used scalar type to vect_get_smallest_scalar_type.
* tree-vect-generic.c (expand_vector_condition): Use
the type of the LHS.
(expand_vector_scalar_condition): Likewise.
(expand_vector_operations_1): Likewise.
* tree-vect-patterns.c (vect_widened_op_tree): Likewise.
(vect_recog_dot_prod_pattern): Likewise.
(vect_recog_sad_pattern): Likewise.
(vect_recog_widen_op_pattern): Likewise.
(vect_recog_widen_sum_pattern): Likewise.
(vect_recog_mixed_size_cond_pattern): Likewise.

Implement EAF_NOT_RETURNED tracking in ipa-modref

2021-07-16  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref.c (struct escape_entry): Use eaf_fleags_t.
(dump_eaf_flags): Dump EAF_NOT_RETURNED
(eaf_flags_useful_p): Use eaf_fleags_t; handle const functions
and EAF_NOT_RETURNED.
(modref_summary::useful_p): Likewise.
(modref_summary_lto::useful_p): Likewise.
(struct) modref_summary_lto: Use eaf_fleags_t.
(deref_flags): Handle EAF_NOT_RETURNED.
(struct escape_point): Use min_flags.
(modref_lattice::init): Add EAF_NOT_RETURNED.
(merge_call_lhs_flags): Ignore EAF_NOT_RETURNED functions
(analyze_ssa_name_flags): Clear EAF_NOT_RETURNED on return;
handle call flags.
(analyze_parms): Also analyze const functions; update conition on
flags usefulness.
(modref_write): Update streaming.
(read_section): Update streaming.
(remap_arg_flags): Use eaf_flags_t.
(modref_merge_call_site_flags): Hanlde EAF_NOT_RETURNED.
* ipa-modref.h: (eaf_flags_t): New typedef.
(struct modref_summary): Use eaf_flags_t.
* tree-core.h (EAF_NOT_RETURNED): New constant.

Get rid of some gimple_expr_type uses

This gets rid of a few gimple_expr_type uses.

2021-07-16  Richard Biener  <rguenther@suse.de>

* gimple-fold.c (gimple_fold_stmt_to_constant_1): Use
the type of the LHS.
(gimple_assign_nonnegative_warnv_p): Likewise.
(gimple_call_nonnegative_warnv_p): Likewise.  Return false
if the call has no LHS.
* gimple.c (gimple_could_trap_p_1): Use the type of the LHS.
* tree-eh.c (stmt_could_throw_1_p): Likewise.
* tree-inline.c (insert_init_stmt): Likewise.
* tree-ssa-loop-niter.c (get_val_for): Likewise.
* tree-outof-ssa.c (ssa_is_replaceable_p): Use the type of
the def.
* tree-ssa-sccvn.c (init_vn_nary_op_from_stmt): Take a
gassign *.  Use the type of the lhs.
(vn_nary_op_lookup_stmt): Adjust.
(vn_nary_op_insert_stmt): Likewise.

IBM Z: Use @PLT symbols for local functions in 64-bit mode

This helps with generating code for kernel hotpatches, which contain
individual functions and are loaded more than 2G away from vmlinux.
This should not create performance regressions for the normal use
cases, because for local functions ld replaces @PLT calls with direct
calls.

gcc/ChangeLog:

* config/s390/predicates.md (bras_sym_operand): Accept all
functions in 64-bit mode, use UNSPEC_PLT31.
(larl_operand): Use UNSPEC_PLT31.
* config/s390/s390.c (s390_loadrelative_operand_p): Likewise.
(legitimize_pic_address): Likewise.
(s390_emit_tls_call_insn): Mark __tls_get_offset as function,
use UNSPEC_PLT31.
(s390_delegitimize_address): Use UNSPEC_PLT31.
(s390_output_addr_const_extra): Likewise.
(print_operand): Add @PLT to TLS calls, handle %K.
(s390_function_profiler): Mark __fentry__/_mcount as function,
use %K, use UNSPEC_PLT31.
(s390_output_mi_thunk): Use only UNSPEC_GOT, use %K.
(s390_emit_call): Use UNSPEC_PLT31.
(s390_emit_tpf_eh_return): Mark __tpf_eh_return as function.
* config/s390/s390.md (UNSPEC_PLT31): Rename from UNSPEC_PLT.
(*movdi_64): Use %K.
(reload_base_64): Likewise.
(*sibcall_brc): Likewise.
(*sibcall_brcl): Likewise.
(*sibcall_value_brc): Likewise.
(*sibcall_value_brcl): Likewise.
(*bras): Likewise.
(*brasl): Likewise.
(*bras_r): Likewise.
(*brasl_r): Likewise.
(*bras_tls): Likewise.
(*brasl_tls): Likewise.
(main_base_64): Likewise.
(reload_base_64): Likewise.
(@split_stack_call<mode>): Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/ext/visibility/noPLT.C: Skip on s390x.
* g++.target/s390/mi-thunk.C: New test.
* gcc.target/s390/nodatarel-1.c: Move foostatic to the new
tests.
* gcc.target/s390/pr80080-4.c: Allow @PLT suffix.
* gcc.target/s390/risbg-ll-3.c: Likewise.
* gcc.target/s390/call.h: Common code for the new tests.
* gcc.target/s390/call-z10-pic-nodatarel.c: New test.
* gcc.target/s390/call-z10-pic.c: New test.
* gcc.target/s390/call-z10.c: New test.
* gcc.target/s390/call-z9-pic-nodatarel.c: New test.
* gcc.target/s390/call-z9-pic.c: New test.
* gcc.target/s390/call-z9.c: New test.
* gcc.target/s390/mfentry-m64-pic.c: New test.
* gcc.target/s390/tls.h: Common code for the new TLS tests.
* gcc.target/s390/tls-pic.c: New test.
* gcc.target/s390/tls.c: New test.

tree-optimization/101467 - fix make_temp_ssa_name usage

My previous change to vect_gen_while introduced paths which call
make_temp_ssa_name with a NULL name which isn't supported.  The
following fixes that.

2021-07-16  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101467
* tree-vect-stmts.c (vect_gen_while): Properly guard
make_temp_ssa_name usage.

testsuite: aarch64: Fix failing SVE tests on big endian

A recent change "gcc: Add vec_select -> subreg RTL simplification"
updated the expected test results for SVE extraction tests. The new
result should only have been changed for little endian. This patch
restores the old expected result for big endian.

gcc/testsuite/ChangeLog:

2021-07-15  Jonathan Wright  <jonathan.wright@arm.com>

* gcc.target/aarch64/sve/extract_1.c: Split expected results
by big/little endian targets, restoring the old expected
result for big endian.
* gcc.target/aarch64/sve/extract_2.c: Likewise.
* gcc.target/aarch64/sve/extract_3.c: Likewise.
* gcc.target/aarch64/sve/extract_4.c: Likewise.

C-SKY: Use the common way to define MULTILIB_DIRNAMES.

C-SKY previously used a forked print-sysroot-suffix.sh and define
CSKY_MULTILIB_DIRNAMES to specify OS multilib directories. This
patch delete the forked print-sysroot-suffix.sh and define
MULTILIB_DIRNAMES to generate same directories.

gcc/
* config.gcc: Don't use forked print-sysroot-suffix.sh and
t-sysroot-suffix for C-SKY.
* config/csky/print-sysroot-suffix.sh: Delete.
* config/csky/t-csky-linux: Delete.
* config/csky/t-sysroot-suffix: Define MULTILIB_DIRNAMES
instead of CSKY_MULTILIB_DIRNAMES.

tree-optimization/101462 - fix signedness of reused reduction vector

This fixes the partial reduction of the reused reduction vector to
carried out in the correct sign and the correctly signed vector
recorded for the skip edge use.

2021-07-16  Richard Biener  <rguenther@suse.de>

* tree-vect-loop.c (vect_transform_cycle_phi): Correct sign
conversion issues with the partial reduction of the reused
vector accumulator.

Disable --param vect-partial-vector-usage by default on x86

The following defaults --param vect-partial-vector-usage to zero
for x86_64 matching existing behavior where support for this
is not present.

2021-07-15  Richard Biener  <rguenther@suse.de>

* config/i386/i386-options.c (ix86_option_override_internal): Set
param_vect_partial_vector_usage to zero if not set.

libstdc++: Adjust doxygen markup for unique_ptr grouping

This reorders the @{ and @relates tags, and moves the definition of the
__cpp_lib_make_unique macro out of the group, as it seems to confuse
doxygen.

libstdc++-v3/ChangeLog:

* include/bits/unique_ptr.h: Adjust doxygen markup.

libstdc++: Adjust doxygen markup for variable templates group [PR101307]

libstdc++-v3/ChangeLog:

PR libstdc++/101307
* include/std/type_traits: Adjust doxygen markup.

libstdc++: Suppress pedantic warnings about __int128

With -std=c++NN -pedantic -Wsystem-headers there are warnings about the
use of __int128, which can be suppressed using diagnostic pragmas.

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

* include/bits/cpp_type_traits.h: Add diagnostic pragmas around
uses of non-standard integer types.
* include/bits/functional_hash.h: Likewise.
* include/bits/iterator_concepts.h: Likewise.
* include/bits/max_size_type.h: Likewise.
* include/bits/std_abs.h: Likewise.
* include/bits/stl_algobase.h: Likewise.
* include/bits/uniform_int_dist.h: Likewise.
* include/ext/numeric_traits.h: Likewise.
* include/std/type_traits: Likewise.

Daily bump.

analyzer: fix const-correctness of various is_a_helper

gcc/analyzer/ChangeLog:
* svalue.h (is_a_helper <placeholder_svalue *>::test): Make
param and template param const.
(is_a_helper <widening_svalue *>::test): Likewise.
(is_a_helper <compound_svalue *>::test): Likewise.
(is_a_helper <conjured_svalue *>::test): Likewise.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

i386: Fix ix86_hard_regno_mode_ok for TDmode on 32bit targets [PR101346]

General regs on 32bit targets do not support 128bit modes,
including TDmode.

gcc/

2021-07-15  Uroš Bizjak  <ubizjak@gmail.com>

PR target/101346
* config/i386/i386.h (VALID_SSE_REG_MODE): Add TDmode.
(VALID_INT_MODE_P): Add SDmode and DDmode.
Add TDmode for TARGET_64BIT.
(VALID_DFP_MODE_P): Remove.
* config/i386/i386.c (ix86_hard_regno_mode_ok):
Do not use VALID_DFP_MODE_P.

gcc/testsuite/

2021-07-15  Uroš Bizjak  <ubizjak@gmail.com>

PR target/101346
* gcc.target/i386/pr101346.c: New test.

Add gimple_range_type for statements.

The existing mechanisms for picking up the type of a statement are
inconsistent with the needs of ranger. Encapsulate all the bits
required to pick up the return type of a statement in one place, and check
whether the type is supported.

* gimple-range-fold.cc (adjust_pointer_diff_expr): Use
gimple_range_type.
(fold_using_range::fold_stmt): Ditto.
(fold_using_range::range_of_range_op): Ditto.
(fold_using_range::range_of_phi): Ditto.
(fold_using_range::range_of_call): Ditto.
(fold_using_range::range_of_builtin_ubsan_call): Ditto.
(fold_using_range::range_of_builtin_call): Ditto.
(fold_using_range::range_of_cond_expr): Ditto.
* gimple-range-fold.h (gimple_range_type): New.

Detect buffer overflow by aggregate and vector stores [PR97027].

Resolves:
PR middle-end/97027 - missing warning on buffer overflow storing a larger scalar into a smaller array

gcc/ChangeLog:

PR middle-end/97027
* tree-ssa-strlen.c (handle_assign): New function.
(maybe_warn_overflow): Add argument.
(nonzero_bytes_for_type): New function.
(count_nonzero_bytes): Handle more tree types.  Call
nonzero_bytes_for_tye.
(count_nonzero_bytes): Handle types.
(handle_store): Handle stores from function calls.
(strlen_check_and_optimize_call): Move code to handle_assign.  Call
it for assignments from function calls.

gcc/testsuite/ChangeLog:

PR middle-end/97027
* gcc.dg/Wstringop-overflow-15.c: Remove an xfail.
* gcc.dg/Wstringop-overflow-47.c: Adjust xfails.
* gcc.dg/torture/pr69170.c: Avoid valid warnings.
* gcc.dg/torture/pr70025.c: Prune out a false positive.
* gcc.dg/vect/pr97769.c: Initialize a loop control variable.
* gcc.target/i386/pr92658-avx512bw-trunc.c: Increase buffer size
to avoid overflow.
* gcc.target/i386/pr92658-avx512f.c: Same.
* gcc.dg/Wstringop-overflow-68.c: New test.
* gcc.dg/Wstringop-overflow-69.c: New test.
* gcc.dg/Wstringop-overflow-70.c: New test.
* gcc.dg/Wstringop-overflow-71.c: New test.
* gcc.dg/strlenopt-95.c: New test.

analyzer: reimplement -Wanalyzer-use-of-uninitialized-value [PR95006 et al]

The initial gcc 10 era commit of the analyzer (in
757bf1dff5e8cee34c0a75d06140ca972bfecfa7) had an implementation of
-Wanalyzer-use-of-uninitialized-value, but was sufficiently buggy
that I removed it in 78b9783774bfd3540f38f5b1e3c7fc9f719653d7 before
the release of gcc 10.1

This patch reintroduces the warning, heavily rewritten, with (I hope)
a less buggy implementation this time, for GCC 12.

gcc/analyzer/ChangeLog:
PR analyzer/95006
PR analyzer/94713
PR analyzer/94714
* analyzer.cc (maybe_reconstruct_from_def_stmt): Split out
GIMPLE_ASSIGN case into...
(get_diagnostic_tree_for_gassign_1): New.
(get_diagnostic_tree_for_gassign): New.
* analyzer.h (get_diagnostic_tree_for_gassign): New decl.
* analyzer.opt (Wanalyzer-write-to-string-literal): New.
* constraint-manager.cc (class svalue_purger): New.
(constraint_manager::purge_state_involving): New.
* constraint-manager.h
(constraint_manager::purge_state_involving): New.
* diagnostic-manager.cc (saved_diagnostic::supercedes_p): New.
(dedupe_winners::handle_interactions): New.
(diagnostic_manager::emit_saved_diagnostics): Call it.
* diagnostic-manager.h (saved_diagnostic::supercedes_p): New decl.
* engine.cc (impl_region_model_context::warn): Convert return type
to bool.  Return false if the diagnostic isn't saved.
(impl_region_model_context::purge_state_involving): New.
(impl_sm_context::get_state): Use NULL ctxt when querying old
rvalue.
(impl_sm_context::set_next_state): Use new sval when querying old
state.
(class dump_path_diagnostic): Move to region-model.cc
(exploded_node::on_stmt): Move to on_stmt_pre and on_stmt_post.
Remove call to purge_state_involving.
(exploded_node::on_stmt_pre): New, based on the above.  Move most
of it to region_model::on_stmt_pre.
(exploded_node::on_stmt_post): Likewise, moving to
region_model::on_stmt_post.
(class stale_jmp_buf): Fix parent class to use curiously recurring
template pattern.
(feasibility_state::maybe_update_for_edge): Call on_call_pre and
on_call_post on gcalls.
* exploded-graph.h (impl_region_model_context::warn): Return bool.
(impl_region_model_context::purge_state_involving): New decl.
(exploded_node::on_stmt_pre): New decl.
(exploded_node::on_stmt_post): New decl.
* pending-diagnostic.h (pending_diagnostic::use_of_uninit_p): New.
(pending_diagnostic::supercedes_p): New.
* program-state.cc (sm_state_map::get_state): Inherit state for
conjured_svalue as well as initial_svalue.
(sm_state_map::purge_state_involving): Also support SK_CONJURED.
* region-model-impl-calls.cc (call_details::get_uncertainty):
Handle m_ctxt being NULL.
(call_details::get_or_create_conjured_svalue): New.
(region_model::impl_call_fgets): New.
(region_model::impl_call_fread): New.
* region-model-manager.cc
(region_model_manager::get_or_create_initial_value): Return an
uninitialized poisoned value for regions that can't have initial
values.
* region-model-reachability.cc
(reachable_regions::mark_escaped_clusters): Handle ctxt being
NULL.
* region-model.cc (region_to_value_map::purge_state_involving): New.
(poisoned_value_diagnostic::use_of_uninit_p): New.
(poisoned_value_diagnostic::emit): Handle POISON_KIND_UNINIT.
(poisoned_value_diagnostic::describe_final_event): Likewise.
(region_model::check_for_poison): New.
(region_model::on_assignment): Call it.
(class dump_path_diagnostic): Move here from engine.cc.
(region_model::on_stmt_pre): New, based on exploded_node::on_stmt.
(region_model::on_call_pre): Move the setting of the LHS to a
conjured svalue to before the checks for specific functions.
Handle "fgets", "fgets_unlocked", and "fread".
(region_model::purge_state_involving): New.
(region_model::handle_unrecognized_call): Handle ctxt being NULL.
(region_model::get_rvalue): Call check_for_poison.
(selftest::test_stack_frames): Use NULL for context when getting
uninitialized rvalue.
(selftest::test_alloca): Likewise.
* region-model.h (region_to_value_map::purge_state_involving): New
decl.
(call_details::get_or_create_conjured_svalue): New decl.
(region_model::on_stmt_pre): New decl.
(region_model::purge_state_involving): New decl.
(region_model::impl_call_fgets): New decl.
(region_model::impl_call_fread): New decl.
(region_model::check_for_poison): New decl.
(region_model_context::warn): Return bool.
(region_model_context::purge_state_involving): New.
(noop_region_model_context::warn): Return bool.
(noop_region_model_context::purge_state_involving): New.
(test_region_model_context:: warn): Return bool.
* region.cc (region::get_memory_space): New.
(region::can_have_initial_svalue_p): New.
(region::involves_p): New.
* region.h (enum memory_space): New.
(region::get_memory_space): New decl.
(region::can_have_initial_svalue_p): New decl.
(region::involves_p): New decl.
* sm-malloc.cc (use_after_free::supercedes_p): New.
* store.cc (binding_cluster::purge_state_involving): New.
(store::purge_state_involving): New.
* store.h (class symbolic_binding): New forward decl.
(binding_key::dyn_cast_symbolic_binding): New.
(symbolic_binding::dyn_cast_symbolic_binding): New.
(binding_cluster::purge_state_involving): New.
(store::purge_state_involving): New.
* svalue.cc (svalue::can_merge_p): Reject attempts to merge
poisoned svalues with other svalues, so that we identify
paths in which a variable is conditionally uninitialized.
(involvement_visitor::visit_conjured_svalue): New.
(svalue::involves_p): Also handle SK_CONJURED.
(poison_kind_to_str): Handle POISON_KIND_UNINIT.
(poisoned_svalue::maybe_fold_bits_within): New.
* svalue.h (enum poison_kind): Add POISON_KIND_UNINIT.
(poisoned_svalue::maybe_fold_bits_within): New decl.

gcc/ChangeLog:
PR analyzer/95006
PR analyzer/94713
PR analyzer/94714
* doc/invoke.texi: Add -Wanalyzer-use-of-uninitialized-value.

gcc/testsuite/ChangeLog:
PR analyzer/95006
PR analyzer/94713
PR analyzer/94714
* g++.dg/analyzer/pr93212.C: Update location of warning.
* g++.dg/analyzer/pr94011.C: Add
-Wno-analyzer-use-of-uninitialized-value.
* g++.dg/analyzer/pr94503.C: Likewise.
* gcc.dg/analyzer/clobbers-1.c: Convert "f" from a local to a
param to avoid uninitialized warning.
* gcc.dg/analyzer/data-model-1.c (test_12): Add test for
uninitialized value on result of alloca.
(test_12a): Add expected warning.
(test_12c): Likewise.
(test_19): Likewise.
(test_29b): Likewise.
(test_29c): Likewise.
(test_37): Remove xfail.
(test_37a): Likewise.
* gcc.dg/analyzer/data-model-20.c: Add warning about leak.
* gcc.dg/analyzer/explode-2.c: Remove params; add
-Wno-analyzer-too-complex, -Wno-analyzer-malloc-leak, and xfails.
Initialize the locals.
* gcc.dg/analyzer/explode-2a.c: Initialize the locals.  Add
expected leak.
* gcc.dg/analyzer/fgets-1.c: New test.
* gcc.dg/analyzer/fread-1.c: New test.
* gcc.dg/analyzer/malloc-1.c (test_16): Add expected warning.
(test_40): Likewise.
* gcc.dg/analyzer/memset-CVE-2017-18549-1.c: Check for
uninitialized padding.
* gcc.dg/analyzer/pr93355-localealias-feasibility.c (fread): New
decl.
(read_alias_file): Call it.
* gcc.dg/analyzer/pr94047.c: Add expected warnings.
* gcc.dg/analyzer/pr94851-2.c: Likewise.
* gcc.dg/analyzer/pr96841.c: Convert local to a param.
* gcc.dg/analyzer/pr98628.c: Likewise.
* gcc.dg/analyzer/pr99042.c: Updated expected location of leak
diagnostics.
* gcc.dg/analyzer/symbolic-1.c: Add expected warnings.
* gcc.dg/analyzer/symbolic-7.c: Likewise.
* gcc.dg/analyzer/torture/pr93649.c: Add expected warning.  Skip
with -fno-fat-lto-objects.
* gcc.dg/analyzer/uninit-1.c: New test.
* gcc.dg/analyzer/uninit-2.c: New test.
* gcc.dg/analyzer/uninit-3.c: New test.
* gcc.dg/analyzer/uninit-4.c: New test.
* gcc.dg/analyzer/uninit-pr94713.c: New test.
* gcc.dg/analyzer/uninit-pr94714.c: New test.
* gcc.dg/analyzer/use-after-free-2.c: New test.
* gcc.dg/analyzer/use-after-free-3.c: New test.
* gcc.dg/analyzer/zlib-3.c: Add expected warning.
* gcc.dg/analyzer/zlib-6.c: Convert locals to params to avoid
uninitialized warnings.  Remove xfail.
* gcc.dg/analyzer/zlib-6a.c: New test, based on the old version
of the above.
* gfortran.dg/analyzer/pr97668.f: Add
-Wno-analyzer-use-of-uninitialized-value and
-Wno-analyzer-too-complex.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

analyzer: add -fdump-analyzer-exploded-paths

gcc/analyzer/ChangeLog:
* analyzer.opt (fdump-analyzer-exploded-paths): New.
* diagnostic-manager.cc
(diagnostic_manager::emit_saved_diagnostic): Implement it.
* engine.cc (exploded_path::dump_to_pp): Add ext_state param and
use it to dump states if non-NULL.
(exploded_path::dump): Likewise.
(exploded_path::dump_to_file): New.
* exploded-graph.h (exploded_path::dump_to_pp): Add ext_state
param.
(exploded_path::dump): Likewise.
(exploded_path::dump): Likewise.
(exploded_path::dump_to_file): New.

gcc/ChangeLog:
* doc/invoke.texi (-fdump-analyzer-exploded-paths): New.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

analyzer: use DECL_DEBUG_EXPR on SSA names for artificial vars

gcc/analyzer/ChangeLog:
* analyzer.cc (fixup_tree_for_diagnostic_1): Use DECL_DEBUG_EXPR
if it's available.
* engine.cc (readability): Likewise.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

analyzer: handle self-referential phis

gcc/analyzer/ChangeLog:
* state-purge.cc (self_referential_phi_p): New.
(state_purge_per_ssa_name::process_point): Don't purge an SSA name
at its def-stmt if the def-stmt is self-referential.

gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/phi-1.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

testsuite: [arm] Remove arm_v8_2a_imm8_neon_ok_nocache

This patch removes this recently-introduced effective-target, as it
looks like a typo and duplicate for arm_v8_2a_i8mm_ok (imm8 vs i8mm),
and it is not used.

2021-07-15  Christophe Lyon  <christophe.lyon@foss.st.com>

gcc/testsuite/
* lib/target-supports.exp (arm_v8_2a_imm8_neon_ok_nocache):
Delete.

testsuite: [arm] Add missing effective-target to vusdot-autovec.c

This test fails when forcing an -mcpu option incompatible with
-march=armv8.2-a+i8mm.

This patch adds the missing arm_v8_2a_i8mm_ok effective-target, as
well as the associated dg-add-options arm_v8_2a_i8mm.

2021-07-15  Christophe Lyon  <christophe.lyon@foss.st.com>

gcc/testsuite/
* gcc.target/arm/simd/vusdot-autovec.c: Use arm_v8_2a_i8mm_ok
effective-target.

c++: Optimize away NULLPTR_TYPE comparisons [PR101443]

Comparisons of NULLPTR_TYPE operands cause all kinds of problems in the
middle-end and in fold-const.c, various optimizations assume that if they
see e.g. a non-equality comparison with one of the operands being
INTEGER_CST and it is not INTEGRAL_TYPE_P (which has TYPE_{MIN,MAX}_VALUE),
they can build_int_cst (type, 1) to find a successor.

The following patch fixes it by making sure they don't appear in the IL,
optimize them away at cp_fold time as all can be folded.

Though, I've just noticed that clang++ rejects the non-equality comparisons
instead, foo () > 0 with
invalid operands to binary expression ('decltype(nullptr)' (aka 'nullptr_t') and 'int')
and foo () > nullptr with
invalid operands to binary expression ('decltype(nullptr)' (aka 'nullptr_t') and 'nullptr_t')

Shall we reject those too, in addition or instead of parts of this patch?
If so, wouldn't this patch be still useful for backports, I bet we don't
want to start reject it on the release branches when we used to accept it.

2021-07-15  Jakub Jelinek  <jakub@redhat.com>

PR c++/101443
* cp-gimplify.c (cp_fold): For comparisons with NULLPTR_TYPE
operands, fold them right away to true or false.

* g++.dg/cpp0x/nullptr46.C: New test.

testsuite: Fix testisms in scalar tests PR101457

These testcases accidentally contain the wrong signs for the expected values
for the scalar code.  The vector code however is correct.

Committed as a trivial fix.

gcc/testsuite/ChangeLog:

PR middle-end/101457
* gcc.dg/vect/vect-reduc-dot-17.c: Fix signs of scalar code.
* gcc.dg/vect/vect-reduc-dot-18.c: Likewise.
* gcc.dg/vect/vect-reduc-dot-22.c: Likewise.
* gcc.dg/vect/vect-reduc-dot-9.c: Likewise.

Avoid -Wvla-parameter for nontrivial bounds [PR97548].

Resolves:
PR c/101289 - bogus -Wvla-paramater warning when using const for vla param
PR c/97548 -  bogus -Wvla-parameter on a bound expression involving a parameter

gcc/c-family/ChangeLog:

PR c/101289
PR c/97548
* c-warn.c (warn_parm_array_mismatch): Use OEP_DECL_NAME.

gcc/c/ChangeLog:

PR c/101289
PR c/97548
* c-decl.c (get_parm_array_spec): Strip nops.

gcc/ChangeLog:

PR c/101289
PR c/97548
* fold-const.c (operand_compare::operand_equal_p): Handle OEP_DECL_NAME.
(operand_compare::verify_hash_value): Same.
* tree-core.h (OEP_DECL_NAME): New.

gcc/testsuite/ChangeLog:

* gcc.dg/Wvla-parameter-12.c: New test.

Change the type of return value of profile_count::value to uint64_t

The field in which profile_count holds the count has 61 bits but the
getter method only returns it as a 32 bit number.  The getter is (and
should be) only used for dumping but even dumps are better when they
do not lie.

gcc/ChangeLog:

2021-07-13  Martin Jambor  <mjambor@suse.cz>

* profile-count.h (profile_count::value): Change the return type to
uint64_t.
* gimple-pretty-print.c (dump_gimple_bb_header): Adjust print
statement.
* tree-cfg.c (dump_function_to_file): Likewise.

libstdc++: Fix std::get<T> for std::tuple [PR101427]

The std::get<T> functions relied on deduction failing if more than one
base class existed for the type T.  However the implementation of Core
DR 2303 (in r11-4693) made deduction succeed (and select the
more-derived base class).

This rewrites the implementation of std::get<T> to explicitly check for
more than one occurrence of T in the tuple elements, making it
ill-formed again. Additionally, the large wall of overload resolution
errors described in PR c++/101460 is avoided by making std::get<T> use
__get_helper<I> directly instead of calling std::get<I>, and by adding a
deleted overload of __get_helper<N> for out-of-range N.

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

PR libstdc++/101427
* include/std/tuple (tuple_element): Improve static_assert text.
(__get_helper): Add deleted overload.
(get<i>(tuple<T...>&&), get<i>(const tuple<T...>&&)): Use
__get_helper directly.
(__get_helper2): Remove.
(__find_uniq_type_in_pack): New constexpr helper function.
(get<T>): Use __find_uniq_type_in_pack and __get_helper instead
of __get_helper2.
* testsuite/20_util/tuple/element_access/get_neg.cc: Adjust
expected errors.
* testsuite/20_util/tuple/element_access/101427.cc: New test.

libstdc++: Add noexcept to __replacement_assert [PR101429]

This results in slightly smaller code when assertions are enabled when
either using Clang (because it adds code to call std::terminate when
potentially-throwing functions are called in a noexcept function) or a
freestanding or non-verbose build (because it doesn't use printf).

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

PR libstdc++/101429
* include/bits/c++config (__replacement_assert): Add noexcept.
[!_GLIBCXX_VERBOSE] (__glibcxx_assert_impl): Use __builtin_trap
instead of __replacement_assert.

rs6000: Don't let swaps pass break multiply low-part (PR101129)

2021-07-15  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
PR target/101129
* config/rs6000/rs6000-p8swap.c (has_part_mult): New.
(rs6000_analyze_swaps): Insns containing a subreg of a mult are
not swappable.

gcc/testsuite/
PR target/101129
* gcc.target/powerpc/pr101129.c: New.

Streamline vect_gen_while

This adjusts the vect_gen_while API to match that of
vect_gen_while_not allowing further patches to generate more
than one stmt for the while case.

2021-07-15  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (vect_gen_while): Match up with
vect_gen_while_not.
* tree-vect-stmts.c (vect_gen_while): Adjust API to that
of vect_gen_while_not.
(vect_gen_while_not): Adjust.
* tree-vect-loop-manip.c (vect_set_loop_controls_directly): Likewise.

Abstract out non_null adjustments in ranger.

There are 4 exact copies of the non-null range adjusting code in the
ranger.  This patch abstracts the functionality into a separate method.

As a follow-up I would like to remove the varying_p check, since I have
seen incoming ranges such as [0, 0xff....ef] which are not varying, but
are not-null.  Removing the varying restriction catches those.

gcc/ChangeLog:

* gimple-range-cache.cc (non_null_ref::adjust_range): New.
(ranger_cache::range_of_def): Call adjust_range.
(ranger_cache::entry_range): Same.
* gimple-range-cache.h (non_null_ref::adjust_range): New.
* gimple-range.cc (gimple_ranger::range_of_expr): Call
adjust_range.
(gimple_ranger::range_on_entry): Same.

Revert "AArch32: Correct sdot RTL on aarch32"

This reverts commit c9165e2d58bb037793c1c93e1b5633a61f88db30.

Revert "AArch64: Correct dot-product auto-vect optab RTL"

This reverts commit 6d1cdb27828d2ef1ae1ab0209836646a269b9610.

gimplify: Fix endless recursion on volatile empty type reads/writes [PR101437]

Andrew's recent change to optimize away during gimplification not just
assignments of zero sized types, but also assignments of empty types,
caused infinite recursion in the gimplifier.
If such assignment is optimized away, we gimplify separately the to_p
and from_p operands and throw away the result.  When gimplifying the
operand that is volatile, we run into the gimplifier code below, which has
different handling for types with non-BLKmode mode, tries to gimplify
those as vol.N = expr, and for BLKmode just throws those away.
Zero sized types will always have BLKmode and so are fine, but for the
non-BLKmode ones like struct S in the testcase, the vol.N = expr
gimplification will reach again the gimplify_modify_expr code, see it is
assignment of empty type and will gimplify again vol.N separately
(non-volatile, so ok) and expr, on which it will recurse again.

The following patch breaks that infinite recursion by ignoring bare
volatile loads from empty types.
If a volatile load or store for aggregates are supposed to be member-wise
loads or stores, then there are no non-padding members in the empty types that
should be copied and so it is probably ok.

2021-07-15  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/101437
* gimplify.c (gimplify_expr): Throw away volatile reads from empty
types even if they have non-BLKmode TYPE_MODE.

* gcc.c-torture/compile/pr101437.c: New test.

[POWER10] __morestack calls from pcrel code

Compiling gcc/testsuite/gcc.dg/split-*.c and others with -mcpu=power10
and linking with a non-pcrel libgcc results in crashes due to the
power10 pcrel code not having r2 set for the generic-morestack.c
functions called from __morestack.  There is also a problem when
non-pcrel code calls a pcrel libgcc.  See the patch comments.

A similar situation theoretically occurs with ELFv1 multi-toc
executables, when __morestack might be located in a different toc
group to its caller.  This patch makes no attempt to fix that, since
the gold linker does not support multi-toc (gold is needed for proper
support of -fsplit-stack code) nor does gcc emit __morestack calls
that support multi-toc.

* config/rs6000/morestack.S (R2_SAVE): Define.
(__morestack): Save and restore r2.  Set up r2 for called
functions.

driver/101383 - handle -gtoggle in driver

The driver amends assembler options with for example --gdwarf-5
when debugging is enabled but the check for that does not consider
the effect of -gtoggle which is not handled in the common option
machinery.  The following alters debug_info_level according to
-gtoggle mimicing what process_options later does in the compiler.

This in particular avoids changing of the cc1-checksum with every
bootstrap (debug) cycle as we compute that from stage2 where we
use -g -gtoggle but with --gdwarf-5 and no debug info from the
compiler the assembler will fill the line table with the temporary
assembler file names.

2021-07-09  Richard Biener  <rguenther@suse.de>

PR driver/101383
* gcc.c (process_command): Process -gtoggle like process_options
would after parsing options.

add myself to DCO section

ChangeLog:

* MAINTAINERS: Add myself to DCO section.

Signed-off-by: Trevor Saunders <tbsaunde@tbsaunde.org>

pass location to md_asm_adjust

So the hook can use it as the location of diagnostics.

gcc/ChangeLog:

* cfgexpand.c (expand_asm_loc): Adjust.
(expand_asm_stmt): Likewise.
* config/arm/aarch-common-protos.h (arm_md_asm_adjust): Likewise.
* config/arm/aarch-common.c (arm_md_asm_adjust): Likewise.
* config/arm/arm.c (thumb1_md_asm_adjust): Likewise.
* config/avr/avr.c (avr_md_asm_adjust): Likewise.
* config/cris/cris.c (cris_md_asm_adjust): Likewise.
* config/i386/i386.c (ix86_md_asm_adjust): Likewise.
* config/mn10300/mn10300.c (mn10300_md_asm_adjust): Likewise.
* config/nds32/nds32.c (nds32_md_asm_adjust): Likewise.
* config/pdp11/pdp11.c (pdp11_md_asm_adjust): Likewise.
* config/rs6000/rs6000.c (rs6000_md_asm_adjust): Likewise.
* config/s390/s390.c (s390_md_asm_adjust): Likewise.
* config/vax/vax.c (vax_md_asm_adjust): Likewise.
* config/visium/visium.c (visium_md_asm_adjust): Likewise.
* doc/tm.texi: Regenerate.
* target.def: Add location argument to md_asm_adjust.

Signed-off-by: Trevor Saunders <tbsaunde@tbsaunde.org>

use diagnostic location in diagnostic_report_current_function

It appears that input_location was used here before the diagnostic's location
was available, and never updated, when the other part of the header was added
that uses it, so this makes it consistent.

gcc/ChangeLog:

* tree-diagnostic.c (diagnostic_report_current_function): Use the
diagnostic's location, not input_location.

Signed-off-by: Trevor Saunders <tbsaunde@tbsaunde.org>

use error_at and warning_at in cfgexpand.c

gcc/ChangeLog:

* cfgexpand.c (tree_conflicts_with_clobbers_p): Pass location to
diagnostics.
(expand_asm_stmt): Likewise.

Signed-off-by: Trevor Saunders <tbsaunde@tbsaunde.org>

c++: fix tree_contains_struct for C++ types [PR101095]

Many of the types from cp-tree.def were only marked as having tree_common,
when actually most of them have type_non_common.  This broke
g++.dg/modules/xtreme-header-2, as the modules code relies on
tree_contains_struct to know what bits it needs to stream.

We don't seem to use type_non_common for TYPE_ARGUMENT_PACK, so I bumped it
down to TS_TYPE_COMMON.  I tried doing the same in cp_tree_size, but that
breaks without more extensive changes to tree_node_structure.

Why do we need the init_ts function anyway?  It seems redundant with
tree_node_structure.

PR c++/101095

gcc/cp/ChangeLog:

* cp-objcp-common.c (cp_common_init_ts): Mark types as types.
(cp_tree_size): Remove redundant entries.

Daily bump.

rs6000: Generate an lxvp instead of two adjacent lxv instructions

The MMA build built-ins currently use individual lxv instructions to
load up the registers of a __vector_pair or __vector_quad.  If the
memory addresses of the built-in operands are to adjacent locations,
then we can use an lxvp in some cases to load up two registers at once.
The patch below adds support for checking whether memory addresses are
adjacent and emitting an lxvp instead of two lxv instructions.

2021-07-14  Peter Bergner  <bergner@linux.ibm.com>

gcc/
* config/rs6000/rs6000.c (adjacent_mem_locations): Return the lower
addressed memory rtx, if any.
(rs6000_split_multireg_move): Fix code formatting.
Handle MMA build built-ins with operands in adjacent memory locations.

gcc/testsuite/
* gcc.target/powerpc/mma-builtin-9.c: New test.

rs6000: Move rs6000_split_multireg_move to later in file

An upcoming change to rs6000_split_multireg_move requires it to be
moved later in the file to fix a declaration issue.

2021-07-14  Peter Bergner  <bergner@linux.ibm.com>

gcc/
* config/rs6000/rs6000.c (rs6000_split_multireg_move): Move to later
in the file.

c++: CTAD and forwarding references [PR88252]

Here during CTAD we're incorrectly treating T&& as a forwarding
reference even though T is a template parameter of the class template.

This happens because the template parameter T in the out-of-line
definition of the constructor doesn't have the flag
TEMPLATE_TYPE_PARM_FOR_CLASS set, and during duplicate_decls the
the redeclaration (which is in terms of this unflagged T) prevails.
To fix this, we could perhaps be more consistent about setting the flag,
but it appears we don't really need this flag to make the determination.

Since the template parameters of an synthesized guide consist of the
template parameters of the class template followed by those of the
constructor (if any), it should suffice to look at the index of the
template parameter to determine whether it comes from the class
template or the constructor (template).  This patch replaces the
TEMPLATE_TYPE_PARM_FOR_CLASS flag with this approach.

PR c++/88252

gcc/cp/ChangeLog:

* cp-tree.h (TEMPLATE_TYPE_PARM_FOR_CLASS): Remove.
* pt.c (push_template_decl): Remove TEMPLATE_TYPE_PARM_FOR_CLASS
handling.
(redeclare_class_template): Likewise.
(forwarding_reference_p): Define.
(maybe_adjust_types_for_deduction): Use it instead.  Add 'tparms'
parameter.
(unify_one_argument): Pass tparms to
maybe_adjust_types_for_deduction.
(try_one_overload): Likewise.
(unify): Likewise.
(rewrite_template_parm): Remove TEMPLATE_TYPE_PARM_FOR_CLASS
handling.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1z/class-deduction96.C: New test.

vec: use auto_vec in a few more places

The uses of vec<T> in get_all_loop_exits and process_conditional were memory
leaks, as .release() was never called for them.  The other changes are some
cases that did have proper release handling, but it's simpler to leave
releasing to the auto_vec destructor.

gcc/ChangeLog:

* sel-sched-ir.h (get_all_loop_exits): Use auto_vec.

gcc/cp/ChangeLog:

* class.c (struct find_final_overrider_data): Use auto_vec.
(find_final_overrider): Remove explicit release.
* coroutines.cc (process_conditional): Use auto_vec.
* cp-gimplify.c (struct cp_genericize_data): Use auto_vec.
(cp_genericize_tree): Remove explicit release.
* parser.c (cp_parser_objc_at_property_declaration): Use
auto_delete_vec.
* semantics.c (omp_reduction_lookup): Use auto_vec.

c++: enable -fdelete-dead-exceptions by default

As I was discussing with richi, I don't think it makes sense to protect
calls to pure/const functions from DCE just because they aren't explicitly
declared noexcept.  PR100382 indicates that there are different
considerations for Go, which has non-call exceptions.  But still turn the
flag off for that specific testcase.

gcc/c-family/ChangeLog:

* c-opts.c (c_common_post_options): Set -fdelete-dead-exceptions.

gcc/ChangeLog:

* doc/invoke.texi: -fdelete-dead-exceptions is on by default for
C++.

gcc/testsuite/ChangeLog:

* g++.dg/torture/pr100382.C: Pass -fno-delete-dead-exceptions.

Vect: correct rebase issue

The lines being removed have been updated and merged into a new
condition.  But when resolving some conflicts I accidentally
reintroduced them causing some test failes.

This removes them.

Committed as the changes were previously approved in
https://gcc.gnu.org/pipermail/gcc-patches/2021-July/574977.html
but the hunk was misapplied during a rebase.

gcc/ChangeLog:

* tree-vect-patterns.c (vect_recog_dot_prod_pattern):
Remove erroneous line.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/vect-reduc-dot-11.c: Expect pass.
* gcc.dg/vect/vect-reduc-dot-15.c: Likewise.
* gcc.dg/vect/vect-reduc-dot-19.c: Likewise.
* gcc.dg/vect/vect-reduc-dot-21.c: Likewise.

Turn hybrid mode off, default to ranger-only mode for EVRP.

Change the default EVRP mode to ranger-only.

gcc/
* params.opt (param_evrp_mode): Change default.

gcc/testsuite/
* gcc.dg/pr80776-1.c: Remove xfail.

c++: constexpr array reference and value-initialization [PR101371]

This PR gave me a hard time: I saw multiple issues starting with
different revisions.  But ultimately the root cause seems to be
the following, and the attached patch fixes all issues I've found
here.

In cxx_eval_array_reference we create a new constexpr context for the
CP_AGGREGATE_TYPE_P case, but we also have to create it for the
non-aggregate case.  In this test, we are evaluating

  ((B *)this)->a = rhs->a

which means that we set ctx.object to ((B *)this)->a.  Then we proceed
to evaluate the initializer, rhs->a.  For *rhs, we eval rhs, a PARM_DECL,
for which we have (const B &) &c.arr[0] in the hash table.  Then
cxx_fold_indirect_ref gives us c.arr[0].  c is evaluated to {.arr={}} so
c.arr is {}.  Now we want c.arr[0], so we end up in cxx_eval_array_reference
and since we're initializing from {}, we call build_value_init which
gives us an AGGR_INIT_EXPR that calls 'constexpr B::B()'.  Then we
evaluate this AGGR_INIT_EXPR and since its first argument is dummy,
we take ctx.object instead.  But that is the wrong object, we're not
initializing ((B *)this)->a here.  And so we wound up with an
initializer for A, and then crash in cxx_eval_component_reference:

  gcc_assert (DECL_CONTEXT (part) == TYPE_MAIN_VARIANT (TREE_TYPE (whole)));

where DECL_CONTEXT (part) is B (as it should be) but the type of whole
was A.

So create a new object, if there already was one, and the element type
is not a scalar.

PR c++/101371

gcc/cp/ChangeLog:

* constexpr.c (cxx_eval_array_reference): Create a new .object
and .ctor for the non-aggregate non-scalar case too when
value-initializing.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1y/constexpr-101371-2.C: New test.
* g++.dg/cpp1y/constexpr-101371.C: New test.

Fortran - ICE in gfc_conv_expr_present initializing non-dummy class variable

gcc/fortran/ChangeLog:

PR fortran/100949
* trans-expr.c (gfc_trans_class_init_assign): Call
gfc_conv_expr_present only for dummy variables.

gcc/testsuite/ChangeLog:

PR fortran/100949
* gfortran.dg/pr100949.f90: New test.

AArch64: Correct dot-product auto-vect optab RTL

The current RTL for the vectorizer patterns for dot-product are incorrect.
Operand3 isn't an output parameter so we can't write to it.

This fixes this issue and reduces the number of RTL.

gcc/ChangeLog:

* config/aarch64/aarch64-simd-builtins.def (udot, sdot): Rename to...
(sdot_prod, udot_prod): ...These.
* config/aarch64/aarch64-simd.md (<sur>dot_prod<vsi2qi>): Remove.
(aarch64_<sur>dot<vsi2qi>): Rename to...
(<sur>dot_prod<vsi2qi>): ...This.
* config/aarch64/arm_neon.h (vdot_u32, vdotq_u32, vdot_s32, vdotq_s32):
Update builtins.

AArch32: Correct sdot RTL on aarch32

The RTL Generated from <sup>dot_prod<vsi2qi> is invalid as operand3 cannot be
written to, it's a normal input.  For the expand it's just another operand
but the caller does not expect it to be written to.

gcc/ChangeLog:

* config/arm/neon.md (<sup>dot_prod<vsi2qi>): Drop statements.

middle-end: Add tests middle end generic tests for sign differing dotproduct.

This adds testcases to test for auto-vect detection of the new sign differing
dot product.

gcc/ChangeLog:

* doc/sourcebuild.texi (arm_v8_2a_i8mm_neon_hw): Document.

gcc/testsuite/ChangeLog:

* lib/target-supports.exp
(check_effective_target_arm_v8_2a_imm8_neon_ok_nocache,
check_effective_target_arm_v8_2a_i8mm_neon_hw,
check_effective_target_vect_usdot_qi): New.
* gcc.dg/vect/vect-reduc-dot-9.c: New test.
* gcc.dg/vect/vect-reduc-dot-10.c: New test.
* gcc.dg/vect/vect-reduc-dot-11.c: New test.
* gcc.dg/vect/vect-reduc-dot-12.c: New test.
* gcc.dg/vect/vect-reduc-dot-13.c: New test.
* gcc.dg/vect/vect-reduc-dot-14.c: New test.
* gcc.dg/vect/vect-reduc-dot-15.c: New test.
* gcc.dg/vect/vect-reduc-dot-16.c: New test.
* gcc.dg/vect/vect-reduc-dot-17.c: New test.
* gcc.dg/vect/vect-reduc-dot-18.c: New test.
* gcc.dg/vect/vect-reduc-dot-19.c: New test.
* gcc.dg/vect/vect-reduc-dot-20.c: New test.
* gcc.dg/vect/vect-reduc-dot-21.c: New test.
* gcc.dg/vect/vect-reduc-dot-22.c: New test.

AArch32: Add support for sign differing dot-product usdot for NEON.

This adds optabs implementing usdot_prod.

The following testcase:

#define N 480
#define SIGNEDNESS_1 unsigned
#define SIGNEDNESS_2 signed
#define SIGNEDNESS_3 signed
#define SIGNEDNESS_4 unsigned

SIGNEDNESS_1 int __attribute__ ((noipa))
f (SIGNEDNESS_1 int res, SIGNEDNESS_3 char *restrict a,
   SIGNEDNESS_4 char *restrict b)
{
  for (__INTPTR_TYPE__ i = 0; i < N; ++i)
    {
      int av = a[i];
      int bv = b[i];
      SIGNEDNESS_2 short mult = av * bv;
      res += mult;
    }
  return res;
}

Generates

f:
        vmov.i32        q8, #0  @ v4si
        add     r3, r2, #480
.L2:
        vld1.8  {q10}, [r2]!
        vld1.8  {q9}, [r1]!
        vusdot.s8       q8, q9, q10
        cmp     r3, r2
        bne     .L2
        vadd.i32        d16, d16, d17
        vpadd.i32       d16, d16, d16
        vmov.32 r3, d16[0]
        add     r0, r0, r3
        bx      lr

instead of

f:
        vmov.i32        q8, #0  @ v4si
        add     r3, r2, #480
.L2:
        vld1.8  {q9}, [r2]!
        vld1.8  {q11}, [r1]!
        cmp     r3, r2
        vmull.s8 q10, d18, d22
        vmull.s8 q9, d19, d23
        vaddw.s16       q8, q8, d20
        vaddw.s16       q8, q8, d21
        vaddw.s16       q8, q8, d18
        vaddw.s16       q8, q8, d19
        bne     .L2
        vadd.i32        d16, d16, d17
        vpadd.i32       d16, d16, d16
        vmov.32 r3, d16[0]
        add     r0, r0, r3
        bx      lr

For NEON.  I couldn't figure out if the MVE instruction vmlaldav.s16 could be
used to emulate this.  Because it would require additional widening to work I
left MVE out of this patch set but perhaps someone should take a look.

gcc/ChangeLog:

* config/arm/neon.md (usdot_prod<vsi2qi>): New.

gcc/testsuite/ChangeLog:

* gcc.target/arm/simd/vusdot-autovec.c: New test.

AArch64: Add support for sign differing dot-product usdot for NEON and SVE.

Hi All,

This adds optabs implementing usdot_prod.

The following testcase:

#define N 480
#define SIGNEDNESS_1 unsigned
#define SIGNEDNESS_2 signed
#define SIGNEDNESS_3 signed
#define SIGNEDNESS_4 unsigned

SIGNEDNESS_1 int __attribute__ ((noipa))
f (SIGNEDNESS_1 int res, SIGNEDNESS_3 char *restrict a,
   SIGNEDNESS_4 char *restrict b)
{
  for (__INTPTR_TYPE__ i = 0; i < N; ++i)
    {
      int av = a[i];
      int bv = b[i];
      SIGNEDNESS_2 short mult = av * bv;
      res += mult;
    }
  return res;
}

Generates for NEON

f:
        movi    v0.4s, 0
        mov     x3, 0
        .p2align 3,,7
.L2:
        ldr     q1, [x2, x3]
        ldr     q2, [x1, x3]
        usdot   v0.4s, v1.16b, v2.16b
        add     x3, x3, 16
        cmp     x3, 480
        bne     .L2
        addv    s0, v0.4s
        fmov    w1, s0
        add     w0, w0, w1
        ret

and for SVE

f:
        mov     x3, 0
        cntb    x5
        mov     w4, 480
        mov     z1.b, #0
        whilelo p0.b, wzr, w4
        mov     z3.b, #0
        ptrue   p1.b, all
        .p2align 3,,7
.L2:
        ld1b    z2.b, p0/z, [x1, x3]
        ld1b    z0.b, p0/z, [x2, x3]
        add     x3, x3, x5
        sel     z0.b, p0, z0.b, z3.b
        whilelo p0.b, w3, w4
        usdot   z1.s, z0.b, z2.b
        b.any   .L2
        uaddv   d0, p1, z1.s
        fmov    x1, d0
        add     w0, w0, w1
        ret

instead of

f:
        movi    v0.4s, 0
        mov     x3, 0
        .p2align 3,,7
.L2:
        ldr     q2, [x1, x3]
        ldr     q1, [x2, x3]
        add     x3, x3, 16
        sxtl    v4.8h, v2.8b
        sxtl2   v3.8h, v2.16b
        uxtl    v2.8h, v1.8b
        uxtl2   v1.8h, v1.16b
        mul     v2.8h, v2.8h, v4.8h
        mul     v1.8h, v1.8h, v3.8h
        saddw   v0.4s, v0.4s, v2.4h
        saddw2  v0.4s, v0.4s, v2.8h
        saddw   v0.4s, v0.4s, v1.4h
        saddw2  v0.4s, v0.4s, v1.8h
        cmp     x3, 480
        bne     .L2
        addv    s0, v0.4s
        fmov    w1, s0
        add     w0, w0, w1
        ret

and

f:
        mov     x3, 0
        cnth    x5
        mov     w4, 480
        mov     z1.b, #0
        whilelo p0.h, wzr, w4
        ptrue   p2.b, all
        .p2align 3,,7
.L2:
        ld1sb   z2.h, p0/z, [x1, x3]
        punpklo p1.h, p0.b
        ld1b    z0.h, p0/z, [x2, x3]
        add     x3, x3, x5
        mul     z0.h, p2/m, z0.h, z2.h
        sunpklo z2.s, z0.h
        sunpkhi z0.s, z0.h
        add     z1.s, p1/m, z1.s, z2.s
        punpkhi p1.h, p0.b
        whilelo p0.h, w3, w4
        add     z1.s, p1/m, z1.s, z0.s
        b.any   .L2
        uaddv   d0, p2, z1.s
        fmov    x1, d0
        add     w0, w0, w1
        ret

gcc/ChangeLog:

* config/aarch64/aarch64-simd.md (aarch64_usdot<vsi2qi>): Rename to...
(usdot_prod<vsi2qi>): ... This.
* config/aarch64/aarch64-simd-builtins.def (usdot): Rename to...
(usdot_prod): ...This.
* config/aarch64/arm_neon.h (vusdot_s32, vusdotq_s32): Likewise.
* config/aarch64/aarch64-sve.md (@aarch64_<sur>dot_prod<vsi2qi>):
Rename to...
(@<sur>dot_prod<vsi2qi>): ...This.
* config/aarch64/aarch64-sve-builtins-base.cc
(svusdot_impl::expand): Use it.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/simd/vusdot-autovec.c: New test.
* gcc.target/aarch64/sve/vusdot-autovec.c: New test.

Vect: Add support for dot-product where the sign for the multiplicant changes.

This patch adds support for a dot product where the sign of the multiplication
arguments differ. i.e. one is signed and one is unsigned but the precisions are
the same.

#define N 480
#define SIGNEDNESS_1 unsigned
#define SIGNEDNESS_2 signed
#define SIGNEDNESS_3 signed
#define SIGNEDNESS_4 unsigned

SIGNEDNESS_1 int __attribute__ ((noipa))
f (SIGNEDNESS_1 int res, SIGNEDNESS_3 char *restrict a,
   SIGNEDNESS_4 char *restrict b)
{
  for (__INTPTR_TYPE__ i = 0; i < N; ++i)
    {
      int av = a[i];
      int bv = b[i];
      SIGNEDNESS_2 short mult = av * bv;
      res += mult;
    }
  return res;
}

The operations are performed as if the operands were extended to a 32-bit value.
As such this operation isn't valid if there is an intermediate conversion to an
unsigned value. i.e.  if SIGNEDNESS_2 is unsigned.

more over if the signs of SIGNEDNESS_3 and SIGNEDNESS_4 are flipped the same
optab is used but the operands are flipped in the optab expansion.

To support this the patch extends the dot-product detection to optionally
ignore operands with different signs and stores this information in the optab
subtype which is now made a bitfield.

The subtype can now additionally controls which optab an EXPR can expand to.

gcc/ChangeLog:

* optabs.def (usdot_prod_optab): New.
* doc/md.texi: Document it and clarify other dot prod optabs.
* optabs-tree.h (enum optab_subtype): Add optab_vector_mixed_sign.
* optabs-tree.c (optab_for_tree_code): Support usdot_prod_optab.
* optabs.c (expand_widen_pattern_expr): Likewise.
* tree-cfg.c (verify_gimple_assign_ternary): Likewise.
* tree-vect-loop.c (vectorizable_reduction): Query dot-product kind.
* tree-vect-patterns.c (vect_supportable_direct_optab_p): Take optional
optab subtype.
(vect_widened_op_tree): Optionally ignore
mismatch types.
(vect_recog_dot_prod_pattern): Support usdot_prod_optab.

x86: Don't enable UINTR in 32-bit mode

UINTR is available only in 64-bit mode.  Since the codegen target is
unknown when the the gcc driver is processing -march=native, to properly
handle UINTR for -march=native:

1. Pass "arch [32|64]" and "tune [32|64]" to host_detect_local_cpu to
indicate 32-bit and 64-bit codegen.
2. Change ix86_option_override_internal to enable UINTR only in 64-bit
mode for -march=CPU when PTA_CPU includes PTA_UINTR.

gcc/

PR target/101395
* config/i386/driver-i386.c (host_detect_local_cpu): Check
"arch [32|64]" and "tune [32|64]" for 32-bit and 64-bit codegen.
Enable UINTR only for 64-bit codegen.
* config/i386/i386-options.c
(ix86_option_override_internal::DEF_PTA): Skip PTA_UINTR if not
in 64-bit mode.
* config/i386/i386.h (ARCH_ARG): New.
(CC1_CPU_SPEC): Pass "[arch|tune] 32" for 32-bit codegen and
"[arch|tune] 64" for 64-bit codegen.

gcc/testsuite/

PR target/101395
* gcc.target/i386/pr101395-1.c: New test.
* gcc.target/i386/pr101395-2.c: Likewise.
* gcc.target/i386/pr101395-3.c: Likewise.

libstdc++: Add noexcept-specifier to basic_string_view(It, End)

This adds a conditional noexcept to the C++20 constructor. The
std::to_address call cannot throw, so only taking the difference of the
two iterators can throw.

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

* include/std/string_view (basic_string_view(It, End)): Add
noexcept-specifier.
* testsuite/21_strings/basic_string_view/cons/char/range.cc:
Check noexcept-specifier. Also check construction without CTAD.

tree-optimization/101445 - fix negative stride SLP vect with gaps

The following fixes the IV adjustment for the gap in a negative
stride SLP vectorization.  The adjustment was in the wrong direction,
now fixes as in the patch.

2021-07-14  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101445
* tree-vect-stmts.c (vectorizable_load): Do the gap adjustment
of the IV in the correct direction for negative stride
accesses.

* gcc.dg/vect/pr101445.c: New testcase.

godump: Fix -fdump-go-spec= reproduceability issue [PR101407]

pot_dummy_types is a hash_set from whose traversal the code prints some type
lines.  hash_set normally uses default_hash_traits which for pointer types
(the hash set hashes const char *) uses pointer_hash which hashes the
addresses of the pointers except of the least significant 3 bits.
With address space randomization, that results in non-determinism in the
-fdump-go-specs= generated file, each invocation can have different order of
the lines emitted from pot_dummy_types traversal.

This patch fixes it by hashing the string contents instead to make the
hashes reproduceable.

2021-07-14  Jakub Jelinek  <jakub@redhat.com>

PR go/101407
* godump.c (godump_str_hash): New type.
(godump_container::pot_dummy_types): Use string_hash instead of
ptr_hash in the hash_set.

Support reduction def re-use for epilogue with different vector size

The following adds support for re-using the vector reduction def
from the main loop in vectorized epilogue loops on architectures
which use different vector sizes for the epilogue.  That's only
x86 as far as I am aware.

2021-07-13  Richard Biener  <rguenther@suse.de>

* tree-vect-loop.c (vect_find_reusable_accumulator): Handle
vector types where the old vector type has a multiple of
the new vector type elements.
(vect_create_partial_epilog): New function, split out from...
(vect_create_epilog_for_reduction): ... here.
(vect_transform_cycle_phi): Reduce the re-used accumulator
to the new vector type.

* gcc.target/i386/vect-reduc-1.c: New testcase.

fix typo in attr_fnspec::verify

Odd-numbered indices describing argument access sizes in the fnspec
string can only hold 't' or a digit, as tested in the beginning of the
case.  When checking that the size-supplying argument does not have
additional information associated with it, the test that excludes the
't' possibility looks for it at the even position in the fnspec
string.  Oops.

This might yield false positives and negatives if a function has a
fnspec in which an argument uses a 't' access-size, and ('t' - '1')
happens to be the index of an argument described in an fnspec string.
Assuming ASCII encoding, it would take a function with at least 68
arguments described in fnspec.  Still, probably worth fixing.

for  gcc/ChangeLog

* tree-ssa-alias.c (attr_fnspec::verify): Fix index in
non-'t'-sized arg check.

adjust landing pads when changing main label

If an artificial label created for a landing pad ends up being
dropped in favor of a user-supplied label, the user-supplied label
inherits the landing pad index, but the post_landing_pad field is not
adjusted to point to the new label.

This patch fixes the problem, and adds verification that we don't
remove a label that's still used as a landing pad.

The circumstance in which this problem can be hit was unusual: removal
of a block with an unreachable label moves the label to some other
unrelated block, in case its address is taken.  In the case at hand
(pr42739.C, complicated by wrappers and cleanups), the chosen block
happened to be an EH landing pad.  (A followup patch will change that.)

for  gcc/ChangeLog

* tree-cfg.c (cleanup_dead_labels_eh): Update
post_landing_pad label upon change of landing pad block's
primary label.
(cleanup_dead_labels): Check that a removed label is not that
of a landing pad.

Daily bump.

gcc: Add vec_select -> subreg RTL simplification

Add a new RTL simplification for the case of a VEC_SELECT selecting
the low part of a vector. The simplification returns a SUBREG.

The primary goal of this patch is to enable better combinations of
Neon RTL patterns - specifically allowing generation of 'write-to-
high-half' narrowing intructions.

Adding this RTL simplification means that the expected results for a
number of tests need to be updated:
* aarch64 Neon: Update the scan-assembler regex for intrinsics tests
  to expect a scalar register instead of lane 0 of a vector.
* aarch64 SVE: Likewise.
* arm MVE: Use lane 1 instead of lane 0 for lane-extraction
  intrinsics tests (as the move instructions get optimized away for
  lane 0.)

This patch also adds new code generation tests to
narrow_high_combine.c to verify the benefit of this RTL
simplification.

gcc/ChangeLog:

2021-06-08  Jonathan Wright  <jonathan.wright@arm.com>

* combine.c (combine_simplify_rtx): Add vec_select -> subreg
simplification.
* config/aarch64/aarch64.md (*zero_extend<SHORT:mode><GPI:mode>2_aarch64):
Add Neon to general purpose register case for zero-extend
pattern.
* config/arm/vfp.md (*arm_movsi_vfp): Remove "*" from *t -> r
case to prevent some cases opting to go through memory.
* cse.c (fold_rtx): Add vec_select -> subreg simplification.
* rtl.c (rtvec_series_p): Define predicate to determine
whether a vector contains a linear series of integers.
* rtl.h (rtvec_series_p): Define.
* rtlanal.c (vec_series_lowpart_p): Define predicate to
determine if a vector selection is equivalent to the low part
of the vector.
* rtlanal.h (vec_series_lowpart_p): Define.
* simplify-rtx.c (simplify_context::simplify_binary_operation_1):
Add vec_select -> subreg simplification.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/extract_zero_extend.c: Remove dump scan
for RTL pattern match.
* gcc.target/aarch64/narrow_high_combine.c: Add new tests.
* gcc.target/aarch64/simd/vmulx_laneq_f64_1.c: Update
scan-assembler regex to look for a scalar register instead of
lane 0 of a vector.
* gcc.target/aarch64/simd/vmulxd_laneq_f64_1.c: Likewise.
* gcc.target/aarch64/simd/vmulxs_lane_f32_1.c: Likewise.
* gcc.target/aarch64/simd/vmulxs_laneq_f32_1.c: Likewise.
* gcc.target/aarch64/simd/vqdmlalh_lane_s16.c: Likewise.
* gcc.target/aarch64/simd/vqdmlals_lane_s32.c: Likewise.
* gcc.target/aarch64/simd/vqdmlslh_lane_s16.c: Likewise.
* gcc.target/aarch64/simd/vqdmlsls_lane_s32.c: Likewise.
* gcc.target/aarch64/simd/vqdmullh_lane_s16.c: Likewise.
* gcc.target/aarch64/simd/vqdmullh_laneq_s16.c: Likewise.
* gcc.target/aarch64/simd/vqdmulls_lane_s32.c: Likewise.
* gcc.target/aarch64/simd/vqdmulls_laneq_s32.c: Likewise.
* gcc.target/aarch64/sve/dup_lane_1.c: Likewise.
* gcc.target/aarch64/sve/extract_1.c: Likewise.
* gcc.target/aarch64/sve/extract_2.c: Likewise.
* gcc.target/aarch64/sve/extract_3.c: Likewise.
* gcc.target/aarch64/sve/extract_4.c: Likewise.
* gcc.target/aarch64/sve/live_1.c: Update scan-assembler regex
cases to look for 'b' and 'h' registers instead of 'w'.
* gcc.target/arm/crypto-vsha1cq_u32.c: Update scan-assembler
regex to reflect lane 0 vector extractions being simplified
to scalar register moves.
* gcc.target/arm/crypto-vsha1h_u32.c: Likewise.
* gcc.target/arm/crypto-vsha1mq_u32.c: Likewise.
* gcc.target/arm/crypto-vsha1pq_u32.c: Likewise.
* gcc.target/arm/mve/intrinsics/vgetq_lane_f16.c: Extract
lane 1 as the moves for lane 0 now get optimized away.
* gcc.target/arm/mve/intrinsics/vgetq_lane_f32.c: Likewise.
* gcc.target/arm/mve/intrinsics/vgetq_lane_s16.c: Likewise.
* gcc.target/arm/mve/intrinsics/vgetq_lane_s32.c: Likewise.
* gcc.target/arm/mve/intrinsics/vgetq_lane_s8.c: Likewise.
* gcc.target/arm/mve/intrinsics/vgetq_lane_u16.c: Likewise.
* gcc.target/arm/mve/intrinsics/vgetq_lane_u32.c: Likewise.
* gcc.target/arm/mve/intrinsics/vgetq_lane_u8.c: Likewise.

rs6000: Add tests for SSE4.1 "test" intrinsics

Copy the test for _mm_testz_si128, _mm_testc_si128,
_mm_testnzc_si128, _mm_test_all_ones, _mm_test_all_zeros,
_mm_test_mix_ones_zeros from gcc/testsuite/gcc.target/i386.

2021-07-13  Paul A. Clarke  <pc@us.ibm.com>

gcc/testsuite
* gcc.target/powerpc/sse4_1-ptest-1.c: Copy from
gcc/testsuite/gcc.target/i386.

rs6000: Add support for SSE4.1 "test" intrinsics

2021-07-13  Paul A. Clarke  <pc@us.ibm.com>

gcc
* config/rs6000/smmintrin.h (_mm_testz_si128, _mm_testc_si128,
_mm_testnzc_si128, _mm_test_all_ones, _mm_test_all_zeros,
_mm_test_mix_ones_zeros): New.

libstdc++: Simplify basic_string_view::ends_with [PR 101361]

The use of npos triggers a diagnostic as described in PR c++/101361.
This change replaces the use of npos with the exact length, which is
already known. We can further simplify it by inlining the effects of
compare and substr, avoiding the redundant range checks in the latter.

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

PR c++/101361
* include/std/string_view (ends_with): Use traits_type::compare
directly.

Adjust testcase to test the call is removed.

Ranger now handles the test.

gcc/testsuite
PR tree-optimization/93781
* gcc.dg/tree-ssa/pr93781-1.c: Check that call is removed.

Make gimple_could_trap_p const-safe.

Allow gimple_could_trap_p (which previously took a non-const gimple)
to be called from functions that take a const gimple (such as
gimple_has_side_effects), and update its prototypes.  Pre-approved
as obvious.

2021-07-13  Roger Sayle  <roger@nextmovesoftware.com>
    Richard Biener  <rguenther@suse.de>

gcc/ChangeLog
* gimple.c (gimple_could_trap_p_1):  Make S argument a
"const gimple*".  Preserve constness in call to
gimple_asm_volatile_p.
(gimple_could_trap_p): Make S argument a "const gimple*".
* gimple.h (gimple_could_trap_p_1, gimple_could_trap_p):
Update function prototypes.

libstdc++: Remove duplicate #include in <string_view>

When I added the new C++23 constructor I added a conditional include of
<bits/ranges_base.h>, which was already being included unconditionally.
This removes the unconditional include but changes the condition for the
other one, so it's used for C++20 as well.

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

* include/std/string_view: Only include <bits/ranges_base.h>
once, and only for C++20 and later.