The class template semiregular-box<T> of [range.semi.wrap] is specified
to value-initialize the underlying object whenever its type is default
initializable. Our primary template for __detail::__box respects this
requirement, but the recently added partial specialization (for types
that are already semiregular) does not.
This patch fixes this issue, and additionally makes the corresponding in
place constructor explicit (as in the primary template).
libstdc++-v3/ChangeLog:
* include/std/ranges (__detail::__box): For the partial
specialization used by types that are already semiregular,
make the default constructor value-initialize the underlying
object instead of default-initializing it. Make its in place
constructor explicit.
* testsuite/std/ranges/adaptors/detail/semiregular_box.cc:
Augment test.
gcc/testsuite/
* g++.dg/guality/guality.exp: Skip $rootme-based check if unset.
* gcc.dg/guality/guality.exp: Likewise.
* gfortran.dg/guality/guality.exp: Likewise.
* lib/asan-dg.exp: Don't use $asan_saved_library_path if not set.
* lib/tsan-dg.exp: Don't use $tsan_saved_library_path if not set.
* lib/ubsan-dg.exp: Don't use $ubsan_saved_library_path if not set.
PR pch/97593
* cgraph.c (cgraph_node::create_thunk): Register thunk as early during
parsing.
* cgraphunit.c (analyze_functions): Call
thunk_info::process_early_thunks.
* symtab-thunks.cc (struct unprocessed_thunk): New struct.
(thunks): New static variable.
(thunk_info::register_early): New member function.
(thunk_info::process_early_thunks): New member function.
* symtab-thunks.h (thunk_info::register_early): Declare.
(thunk_info::process_early_thunks): Declare.
This avoids requiring insert iteration for back-to-back hoisting
opportunities as seen in the added testcase. For the PR at hand
this halves the number of insert iterations retaining only
the hard to avoid PRE / hoist insert back-to-backs.
2020-10-30 Richard Biener <rguenther@suse.de>
PR tree-optimization/97623
* tree-ssa-pre.c (insert): First do hoist insertion in
a backward walk.
* gcc.dg/tree-ssa/ssa-hoist-7.c: New testcase.
This makes sure to roll-back the whole SCC when we fail stmt
analysis, otherwise the optimistic visited treatment breaks down
with different entries. Rollback is easy when tracking additions
to visited in a vector which also makes the whole thing cheaper
than the two hash-sets used before.
2020-10-30 Richard Biener <rguenther@suse.de>
PR tree-optimization/97626
* tree-vect-slp.c (vect_slp_analyze_node_operations):
Exchange the lvisited hash-set for a vector, roll back
recursive adds to visited when analysis failed.
(vect_slp_analyze_operations): Likewise.
* gcc.dg/vect/bb-slp-pr97626.c: New testcase.
After PR92793 commit 9c81750c5b "Fortran] PR
92793 - fix column used for error diagnostic", commit
d0d0ba20f2 did "Add tests to verify OpenACC
clause locations", later fixed up in PR92901 commit
e6c90dba73 to "Fix PR92901: Change test
expectation for C++ in OpenACC test clause-locations.c".
Now, add some more testing to verify/document the status quo.
gcc/testsuite/
PR fortran/92793
* c-c++-common/goacc/clause-locations.c: Rewrite into...
* c-c++-common/goacc/pr92793-1.c: ... this.
* gfortran.dg/goacc/clause-locations.f90: Rewrite into...
* gfortran.dg/goacc/pr92793-1.f90: ... this.
This makes sure to update backedges in single-node cycles.
2020-10-30 Richard Biener <rguenther@suse.de>
PR tree-optimization/97633
* tree-vect-slp.c (): Update backedges in single-node cycles.
Optimize processing of externals.
* g++.dg/vect/slp-pr97636.cc: New testcase.
* gcc.dg/vect/bb-slp-pr97633.c: Likewise.
This change fixes a bug in the AArch64 backend. Currently, we accept an
odd sign_extract representation of addresses, but don't accept that same
odd form of address as an LEA.
This is the cause of PR96998. In the testcase given in the PR, combine
produces:
(insn 9 8 10 3 (set (mem:SI (plus:DI (sign_extract:DI (mult:DI (subreg:DI (reg/v:SI 92 [ g ]) 0)
(const_int 4 [0x4]))
(const_int 34 [0x22])
(const_int 0 [0]))
(reg/f:DI 96)) [3 *i_5+0 S4 A32])
(asm_operands:SI ("") ("=Q") 0 []
[]
[] test.c:11)) "test.c":11:5 -1
(expr_list:REG_DEAD (reg/v:SI 92 [ g ])
(nil)))
Then LRA reloads the address and we ICE because we fail to recognize the
sign_extract outside the mem:
(insn 33 8 34 3 (set (reg:DI 100)
(sign_extract:DI (ashift:DI (subreg:DI (reg/v:SI 92 [ g ]) 0)
(const_int 2 [0x2]))
(const_int 34 [0x22])
(const_int 0 [0]))) "test.c":11:5 -1
(nil))
The aarch64 changes here remove the support for this sign_extract
representation of addresses, fixing PR96998. Now this by itself would
regress code quality, so this change is paired with an improvement to
combine which prevents an extract rtx from being emitted in this case:
we now write the rtx above as a shift of an extend, which allows the
combination to go ahead.
Prior to this, combine.c:make_extraction() identified where we can emit
an ashift of an extend in place of an extraction, but failed to make the
corresponding canonicalization/simplification when presented with a mult
by a power of two. Such a representation is canonical when representing
a left-shifted address inside a mem.
This change remedies this situation. For rtxes such as:
(mult:DI (subreg:DI (reg:SI r) 0) (const_int 2^n))
where the bottom 32 + n bits are valid (the higher-order bits are
undefined) and make_extraction() is being asked to sign_extract the
lower (valid) bits, after the patch, we rewrite this as:
(mult:DI (sign_extend:DI (reg:SI r)) (const_int 2^n))
instead of using a sign_extract.
gcc/ChangeLog:
PR target/96998
* combine.c (make_extraction): Also handle shifts written as
(mult x 2^n), avoid creating an extract rtx for these.
* config/aarch64/aarch64.c (aarch64_is_extend_from_extract): Delete.
(aarch64_classify_index): Remove extract-based address handling.
(aarch64_strip_extend): Likewise.
(aarch64_rtx_arith_op_extract_p): Likewise, remove now-unused parameter.
Update callers...
(aarch64_rtx_costs): ... here.
gcc/testsuite/ChangeLog:
PR target/96998
* gcc.c-torture/compile/pr96998.c: New test.
Unfortunately, some VxWorks 7r2 system headers rely on a
couple more variations of the predefined macros expected
to characterize a "powerpc" target that we discussed recently.
setjmp.h, for example, relies on __ppc and the absence of
a definition results in "gcc" dejagnu test failures from all
the tests #including that header, which stumble on:
#error "_JBLEN not set!"
The other case is __ppc__ expected by yvals.h, key to libstdc++.
This change adjusts the VxWorks 7 section of our configuration
to honor those expectations.
2020-10-29 Olivier Hainque <hainque@adacore.com>
gcc/
* config/rs6000/vxworks.h (TARGET_OS_CPP_BUILTINS): Also
builtin_define __ppc and __ppc__ for VxWorks 7.
This change extends the VxWorks support on intel CPUs to
VxWorks7r2 for x86_64 as well as x86, with a "mcmodel=large"
additional multilib for the 64bit configuration.
The support for fPIC is not functional yet for this model,
so we just don't add the corresponding multilib.
We extend the range of CPU families handled by TARGET_OS_CPP_BUILTINS,
accounting for the fact that archs older than PENTIUM4 are
not supported (any more) by VxWorks 7.
As we did for powerpc, we leverage VX_CPU_PREFIX to emit different
forms of definitions for different families of VxWorks as the system
headers's expectations has evolved between Vx 5, 6 and 7.
2020-10-27 Olivier Hainque <hainque@adacore.com>
gcc/
* config.gcc: Adjust the ix86/x86_64-wrs-vxworks filters
to apply to VxWorks 7 as well.
* config/i386/t-vxworks (MULTILIB_OPTIONS, MULTILIB_DIRNAMES):
Remove the fPIC multilib and add one for the large code model
on x86_64.
* config/i386/vxworks.h: Separate sections for TARGET_VXWORKS7,
other variants and common bits.
(TARGET_OS_CPP_BUILTINS): Augment to support a range of CPU
families. Leverage VX_CPU_PREFIX.
(CC1_SPEC): Add definition.
(STACK_CHECK_PROTECT): Use conditional expression instead of
heavier to read conditioned macro definitions.
libgcc/
* config.host: Adjust the ix86/x86_64-wrs-vxworks filters
to apply to VxWorks 7 as well.
Co-authored-by: Douglas Rupp <rupp@adacore.com>
Co-authored-by: Pat Bernardi <bernardi@adacore.com>
This patch fixes gimplification of allocate clause on taskloop - puts
allocate on inner taskloop only if there is allocate clause, because otherwise
the data sharing clauses are only on the task construct in the construct sandwich.
2020-10-30 Jakub Jelinek <jakub@redhat.com>
* gimplify.c (gimplify_scan_omp_clauses): Force
OMP_CLAUSE_ALLOCATE_ALLOCATOR into a temporary if it is non-NULL and
non-constant.
(gimplify_omp_for): Only put allocate on inner taskloop if lastprivate
for the same variable is going to be put there, and in that case
if the OMP_CLAUSE_ALLOCATE_ALLOCATOR is non-NULL non-constant, make
the allocator firstprivate on task.
* c-c++-common/gomp/allocate-3.c: New test.
This allows specification of non-static data members in allocate clause like it
can be specified in other privatization clauses and adds a new testcase that covers
also handling of that clause in templates.
2020-10-30 Jakub Jelinek <jakub@redhat.com>
* semantics.c (finish_omp_clauses) <case OMP_CLAUSE_ALLOCATE>: Handle
non-static members in methods.
* pt.c (tsubst_omp_clauses): Handle OMP_CLAUSE_ALLOCATE.
* c-c++-common/gomp/allocate-1.c (qux): Add another test.
* g++.dg/gomp/allocate-1.C: New test.
This adjusts the testcase to allow splitting up the group for
larger vector sizes and thus printing the splat message multiple times.
2020-10-30 Richard Biener <rguenther@suse.de>
* gcc.dg/vect/bb-slp-pr65935.c: Adjust.
This patch ccombines two patches:
1) If GLIBC is 2.32 or newer, and the language is C or C++, allow the user to
change the long double type without having to use -Wno-psabi.
2) Adjust the warnings for intermixing __float128 and __ibm128 to accomidate
the future change to allow long double to use the IEEE 128-bit format.
gcc/
2020-10-29 Michael Meissner <meissner@linux.ibm.com>
* config/rs6000/rs6000.c (rs6000_option_override_internal): Allow
long double type to be changed for C/C++ if glibc 2.32 or newer.
(rs6000_invalid_binary_op): Update error messages about mixing IBM
long double and IEEE 128-bit.
gcc/testsuite/
2020-10-27 Michael Meissner <meissner@linux.ibm.com>
* gcc.target/powerpc/float128-mix-2.c: New test.
* gcc.target/powerpc/float128-mix-3.c: New test.
* gcc.target/powerpc/float128-mix.c: Update failure messages.
The 20_util/unique_ptr/creation/for_overwrite.cc testcase relies on
operator new, which requires special features on AIX. This patch
disables the testcase.
libstdc++-v3/ChangeLog:
* testsuite/20_util/unique_ptr/creation/for_overwrite.cc: XFAIL on AIX.
This should fix a bootstrap error on Solaris, due to some of the new
symbols matching old patterns as well as new ones.
libstdc++-v3/ChangeLog:
* config/abi/pre/gnu.ver (GLIBCXX_3.4.21): Tighten patterns
for old <sstream> symbols some more.
libstdc++-v3/ChangeLog:
* include/bits/ranges_util.h (subrange::subrange(R&&)): Use
direct-initialization instead of list-initialization, so a
potential narrowing conversion from ranges::size(r) to the
stored size isn't ill-formed.
These are usually suppressed in system headers, but should be fixed
anyway.
libstdc++-v3/ChangeLog:
* include/bits/parse_numbers.h (_Select_int_base): Avoid
narrowing conversion in constant expression.
* include/experimental/buffer (buffer_copy): Avoid narrowing
conversion.
* include/experimental/internet (hash<>::operator()): Do not
use deprecated 'argument_type' member.
* include/std/variant (variant::emplace): Use cast instead
of implicit conversion from size_t to narrower unsigned type.
This patch implements CWG 625 which prohibits using auto in a template
argument. A few tests used this construction. Since this usage was
allowed by the Concepts TS, we only give an error in C++20.
gcc/cp/ChangeLog:
DR 625
PR c++/97479
* parser.c (cp_parser_type_id_1): Reject using auto as
a template-argument in C++20.
gcc/testsuite/ChangeLog:
DR 625
PR c++/97479
* g++.dg/cpp0x/auto3.C: Update dg-error.
* g++.dg/cpp0x/auto9.C: Likewise.
* g++.dg/cpp2a/concepts-pr84979-2.C: Likewise.
* g++.dg/cpp2a/concepts-pr84979-3.C: Likewise.
* g++.dg/cpp2a/concepts-pr84979.C: Likewise.
* g++.dg/DRs/dr625.C: New test.
In this testcase we weren't able to deduce b's type:
template<typename T> void Task() { }
auto b = { &Task<int> };
because resolve_nondeduced_context doesn't iterate on the {}'s elements.
So make sure to look into {} too. We don't need to handle nested {}
here.
We could either tweak resolve_nondeduced_context to handle CONSTRUCTORs
or add a _ctor version, but then resolve_nondeduced_context_or_error
would need some changes too -- it'd have to check the result of a call
to r_n_c for each element.
gcc/cp/ChangeLog:
PR c++/93107
* pt.c (do_auto_deduction): Call resolve_nondeduced_context for
the elements of a { } list.
gcc/testsuite/ChangeLog:
PR c++/93107
* g++.dg/cpp0x/initlist-deduce3.C: New test.
As [depr.arith.conv.enum] says, these are ill-formed.
gcc/cp/ChangeLog:
* typeck.c (do_warn_enum_conversions): Don't warn for SPACESHIP_EXPR.
(cp_build_binary_op): Reject float <=> enum or enum <=> float. Use
CP_INTEGRAL_TYPE_P instead of INTEGRAL_OR_ENUMERATION_TYPE_P.
gcc/testsuite/ChangeLog:
* g++.dg/cpp2a/enum-conv1.C: Remove unused code.
* g++.dg/cpp2a/spaceship-err5.C: New test.
Many of the high-level constraint normalization routines allow the
caller to supply the initial template arguments for normalization, but
in practice all of the callers supply something equivalent to the
identity mapping(*).
This patch hard-codes this prevalent choice of initial template
arguments by making get_normalized_constraints always pass NULL_TREE as
the args to normalize_expression. This admits some simplifications in
the high-level routines, such as removing their 'args' parameter and
consolidating the two versions of normalize_constraint_expression.
(*): In particular, a set of generic template arguments or NULL_TREE.
In the case of the two-parm version of normalize_constraint_expression,
we were suspiciously using the template arguments of a concept-id when
normalizing the concept-id as a constraint-expression.
gcc/cp/ChangeLog:
* constraint.cc (get_normalized_constraints): Remove 'args'
parameter. Pass NULL_TREE as the initial template arguments to
normalize_expression.
(get_normalized_constraints_from_info): Remove 'args' parameter
and adjust the call to get_normalized_constraints.
(get_normalized_constraints_from_decl): Remove 'args' local
variable and adjust call to get_normalized_constraints_from_info.
(normalize_concept_definition): Remove 'args' local variable
and adjust call to get_normalized_constraints.
(normalize_constraint_expression): Remove the two-parameter
overload. Remove 'args' parameter from the three-parameter
overload and update function comment accordingly. Remove
default argument from 'diag' parameter. Adjust call to
get_normalized_constraints.
(finish_nested_requirement): Adjust call to
normalize_constraint_expression.
(strictly_subsumes): Remove 'args' parameter. Adjust call to
get_normalized_constraints_from_info.
(weakly_subsumes): Likewise.
* cp-tree.h (strictly_subsumes): Remove 'args' parameter.
(weakly_subsumes): Likewise.
* pt.c (process_partial_specialization): Adjust call to
strictly_subsumes.
(is_compatible_template_arg): Adjust call to weakly_subsumes.
When normalizing the constraint-expression of a nested-requirement, we
pass NULL_TREE as the initial template arguments for normalization, but
tsubst_argument_pack is not prepared to handle a NULL_TREE args vector.
This causes us to ICE when normalizing a variadic concept as part of a
nested-requirement.
This patch fixes the ICE by guarding the call to tsubst_template_args in
normalize_concept_check appropriately. This will also enable us to
simplify many of the normalization routines to just pass NULL_TREE
(instead of a set of generic template arguments) as the initial template
arguments.
gcc/cp/ChangeLog:
PR c++/97412
* constraint.cc (normalize_concept_check): Don't call
tsubst_template_args when 'args' is NULL.
gcc/testsuite/ChangeLog:
PR c++/97412
* g++.dg/cpp2a/concepts-variadic2.C: New test.
This fixes some memleaks, one older, one recently introduced.
2020-10-29 Richard Biener <rguenther@suse.de>
* tree-ssa-pre.c (compute_avail): Free operands consistently.
* tree-vect-loop.c (vectorizable_phi): Make sure all operand
defs vectors are released.
In this testcase, the primary evaluation successfully produces 'true', and
then running one of the cleanups hits a double delete, making the whole
thing not a valid constant expression. So we were returning 'true' wrapped
in a NOP_EXPR to indicate its non-constancy, but evaluating that again is a
perfectly acceptable constant expression, so we weren't getting the verbose
diagnostic we were looking for.
So if non_constant_p gets set other than for overflow, go back to the
original expression.
With this change, we should never hit the manifestly_const_eval test, and
the is-constant-evaluated1.C test passes without it.
gcc/cp/ChangeLog:
PR c++/97388
* constexpr.c (cxx_eval_outermost_constant_expr): Revert to
original expression if evaluation sets non_constant_p.
gcc/testsuite/ChangeLog:
PR c++/97388
* g++.dg/cpp2a/constexpr-dtor8.C: New test.
For arguments passed by invisible reference, in the IL until genericization
we have the source types on the callee side and while on the caller side
we already pass references to the actual argument slot in the caller, we
undo that in cxx_bind_parameters_in_call's
if (TREE_ADDRESSABLE (type))
/* Undo convert_for_arg_passing work here. */
x = convert_from_reference (x);
This works fine most of the time, except when the type also has constexpr
destructor; in that case the destructor is invoked in the caller and thus
the unsharing we do to make sure that the callee doesn't modify caller's
values is in that case undesirable, it prevents the changes done in the
callee propagating to the caller which should see them for the constexpr
dtor evaluation.
The following patch fixes that. While it could be perhaps done for all
TREE_ADDRESSABLE types, I don't see the need to change the behavior
if there is no constexpr non-trivial dtor.
Jason: And we need to avoid memoizing the call, because a later equivalent
call also needs to modify its argument. And we don't need to unshare
constructors when we aren't memoizing the call, because we already unshared
them when evaluating the TARGET_EXPR representing the copy-initialization of
the argument.
2020-10-20 Jakub Jelinek <jakub@redhat.com>
Jason Merrill <jason@redhat.com>
PR c++/97388
* constexpr.c (cxx_bind_parameters_in_call): Set non_constant_args
if the parameter type has a non-trivial destructor.
(cxx_eval_call_expression): Only unshare arguments if we're
memoizing this evaluation.
* g++.dg/cpp2a/constexpr-dtor5.C: New test.
* g++.dg/cpp2a/constexpr-dtor6.C: New test.
* g++.dg/cpp2a/constexpr-dtor7.C: New test.
* wide-int.h (trailing_wide_ints <N>): Turn len to array of structures
so it does not imply typeless storage.
(trailing_wide_ints <N>::operator): update
(trailing_wide_ints <N>::operator []): Update.
This patch diagnoses delete [] new int; and delete new int[1]; in constexpr
contexts by remembering
IDENTIFIER_OVL_OP_FLAGS (DECL_NAME (fun)) & OVL_OP_FLAG_VEC
from the operator new and checking it at operator delete time.
2020-10-29 Jakub Jelinek <jakub@redhat.com>
PR c++/95808
* cp-tree.h (enum cp_tree_index): Add CPTI_HEAP_VEC_UNINIT_IDENTIFIER
and CPTI_HEAP_VEC_IDENTIFIER.
(heap_vec_uninit_identifier, heap_vec_identifier): Define.
* decl.c (initialize_predefined_identifiers): Initialize those
identifiers.
* constexpr.c (cxx_eval_call_expression): Reject array allocations
deallocated with non-array deallocation or non-array allocations
deallocated with array deallocation.
(non_const_var_error): Handle heap_vec_uninit_identifier and
heap_vec_identifier too.
(cxx_eval_constant_expression): Handle also heap_vec_uninit_identifier
and in that case during initialization replace it with
heap_vec_identifier.
(find_heap_var_refs): Handle heap_vec_uninit_identifier and
heap_vec_identifier too.
* g++.dg/cpp2a/constexpr-new15.C: New test.
C2x has changed the expansions of the true and false macros in
<stdbool.h> so that they have type _Bool (including in #if conditions,
i.e. an unsigned type in that context). Use the new expansions in
GCC's <stdbool.h> for C2x.
See bug 82272 for related discussion (but this patch does *not*
implement the warning discussed there).
Note that it's possible there may be a further change to make bool,
true and false keywords (there was support in principle for that at
the April WG14 meeting). But currently these expansions of type _Bool
are what C2x requires and there isn't actually a paper before WG14 at
present that would introduce the new keywords.
Bootstrapped with no regressions on x86_64-pc-linux-gnu.
gcc/
2020-10-29 Joseph Myers <joseph@codesourcery.com>
* ginclude/stdbool.h [__STDC_VERSION__ > 201710L] (true, false):
Define with type _Bool.
gcc/testsuite/
2020-10-29 Joseph Myers <joseph@codesourcery.com>
* gcc.dg/c11-bool-1.c, gcc.dg/c2x-bool-1.c, gcc.dg/c99-bool-4.c:
New tests.
These tests just return true without checking that the results of the
algorithms. Although it should be safe to assume that the algorithms
behave the same at compile-time as at run-time, we can use these tests
to verify it.
This replaces each 'return true' statement with a condition that depends
on the basic functionality of the algorithm, such as returning an
iterator to the right position.
libstdc++-v3/ChangeLog:
* testsuite/25_algorithms/all_of/constexpr.cc: Check result of
the algorithm.
* testsuite/25_algorithms/any_of/constexpr.cc: Likewise.
* testsuite/25_algorithms/binary_search/constexpr.cc: Likewise.
* testsuite/25_algorithms/copy_backward/constexpr.cc: Likewise.
* testsuite/25_algorithms/count/constexpr.cc: Likewise.
* testsuite/25_algorithms/equal/constexpr.cc: Likewise.
* testsuite/25_algorithms/equal_range/constexpr.cc: Likewise.
* testsuite/25_algorithms/fill/constexpr.cc: Likewise.
* testsuite/25_algorithms/find_end/constexpr.cc: Likewise.
* testsuite/25_algorithms/find_if/constexpr.cc: Likewise.
* testsuite/25_algorithms/is_partitioned/constexpr.cc: Likewise.
* testsuite/25_algorithms/is_permutation/constexpr.cc: Likewise.
* testsuite/25_algorithms/is_sorted_until/constexpr.cc:
Likewise.
* testsuite/25_algorithms/lexicographical_compare/constexpr.cc:
Likewise.
* testsuite/25_algorithms/lower_bound/constexpr.cc: Likewise.
* testsuite/25_algorithms/merge/constexpr.cc: Likewise.
* testsuite/25_algorithms/mismatch/constexpr.cc: Likewise.
* testsuite/25_algorithms/none_of/constexpr.cc: Likewise.
* testsuite/25_algorithms/partition_copy/constexpr.cc: Likewise.
* testsuite/25_algorithms/remove_copy/constexpr.cc: Likewise.
* testsuite/25_algorithms/remove_copy_if/constexpr.cc: Likewise.
* testsuite/25_algorithms/remove_if/constexpr.cc: Likewise.
* testsuite/25_algorithms/replace_if/constexpr.cc: Likewise.
* testsuite/25_algorithms/reverse/constexpr.cc: Likewise.
* testsuite/25_algorithms/reverse_copy/constexpr.cc: Likewise.
* testsuite/25_algorithms/rotate_copy/constexpr.cc: Likewise.
* testsuite/25_algorithms/search/constexpr.cc: Likewise.
* testsuite/25_algorithms/set_difference/constexpr.cc: Likewise.
* testsuite/25_algorithms/set_intersection/constexpr.cc:
Likewise.
* testsuite/25_algorithms/set_symmetric_difference/constexpr.cc:
Likewise.
* testsuite/25_algorithms/set_union/constexpr.cc: Likewise.
* testsuite/25_algorithms/unique_copy/constexpr.cc: Likewise.
* testsuite/25_algorithms/upper_bound/constexpr.cc: Likewise.
This extends the fast path to also work when the URBG's range of
possible values is not the entire range of its result_type. Previously,
the slow path would be used for engines with a uint_fast32_t result type
if that type is actually a typedef for uint64_t rather than uint32_t.
After this change, the generator's result_type is not important, only
the range of possible value that generator can produce. If the
generator's range is exactly UINT64_MAX then the calculation will be
done using 128-bit and 64-bit integers, and if the range is UINT32_MAX
it will be done using 64-bit and 32-bit integers.
In practice, this benefits most of the engines and engine adaptors
defined in [rand.predef] on x86_64-linux and other 64-bit targets. This
is because std::minstd_rand0 and std::mt19937 and others use
uint_fast32_t, which is a typedef for uint64_t.
The code now makes use of the recently-clarified requirement that the
generator's min() and max() functions are usable in constant
expressions (see LWG 2154).
libstdc++-v3/ChangeLog:
* include/bits/uniform_int_dist.h (_Power_of_two): Add
constexpr.
(uniform_int_distribution::_S_nd): Add static_assert to ensure
the wider type is twice as wide as the result type.
(uniform_int_distribution::__generate_impl): Add static_assert
and declare variables as constexpr where appropriate.
(uniform_int_distribution:operator()): Likewise. Only consider
the uniform random bit generator's range of possible results
when deciding whether _S_nd can be used, not the __uctype type.
The rope extension uses a volatile variable for its reference count.
This is not only unnecessary for correctness (volatile provides neither
atomicity nor memory visibility, and the variable is only modified while
a lock is held) but it now causes deprecated warnings with
-Wsystem-headers due to the use of ++ and -- operators.
It would be possible to use __gnu_cxx::__exchange_and_add in _M_incr and
_M_decr when __atomic_is_lock_free(sizeof(_RC_t), &_M_ref_count) is
true, rather than locking a mutex. That would probably be a significant
improvement for multi-threaded and single-threaded code (because
__exchange_and_add will use non-atomic ops when possible, and even in MT
code it should be faster than the mutex lock/unlock pair). However,
mixing objects compiled with the old and new code would result in
inconsistent synchronization being used for the reference count.
libstdc++-v3/ChangeLog:
* include/ext/rope (_Refcount_Base::_M_ref_count): Remove
volatile qualifier.
(_Refcount_Base::_M_decr()): Likewise.
This change allows std::function::target<F>() to work even without RTTI,
using the same approach as std::any. Because we know what the manager
function would be for a given type, we can check if the stored pointer
has the expected address. If it does, we don't need to use RTTI. If it
isn't equal, we still need to do the RTTI check (when RTTI is enabled)
to handle the case where the same function has different addresses in
different shared objects.
This also changes the implementation of the manager function to return a
null pointer result when asked for the type_info of the target object.
This not only avoids a warning with -Wswitch -Wsystem-headers, but also
avoids prevents std::function::target_type() from dereferencing an
uninitialized pointer when the linker keeps an instantiation of the
manager function that was compiled without RTTI.
Finally, this fixes a bug in the non-const overload of function::target
where calling it with a function type F was ill-formed, due to
attempting to use const_cast<F*>(ptr). The standard only allows
const_cast<T*> when T is an object type. The solution is to use
*const_cast<F**>(&ptr) instead, because F* is an object type even if F
isn't. I've also used _GLIBCXX17_CONSTEXPR in function::target so that
it doesn't bother instantiating anything for types that can never be a
valid target.
libstdc++-v3/ChangeLog:
* include/bits/std_function.h (_Function_handler<void, void>):
Define explicit specialization used for invalid target types.
(_Base_manager::_M_manager) [!__cpp_rtti]: Return null.
(function::target_type()): Check for null pointer.
(function::target()): Define unconditionall. Fix bug with
const_cast of function pointer type.
(function::target() const): Define unconditionally, but
only use RTTI if enabled.
* testsuite/20_util/function/target_no_rtti.cc: New test.
The UBSAN builtins degrade into PLUS/MINUS/MULT and call
extract_range_from_binary_expr, which as the PR shows, can special
case some symbolics which the ranger doesn't currently handle.
Looking at vr_values::extract_range_builtin(), I see that every single
place where we ask for a range, we bail on non-integers (symbolics,
etc). That is, with the exception of the UBSAN builtins.
Since this seems to be particular to UBSAN, we could still go with the
original plan of removing the duplicity in ranger vs vr-values, but
leave in the UBSAN builtin handling. This isn't ideal, as we'd like
to remove all the common code, but I'd be willing to put up with UBSAN
duplication for the time being.
This patch disables the assert on the UBSAN builtins, while still
trapping if any other differences are found between the vr_values and
the ranger versions of builtin range handling.
As a follow-up, once Fedora can test this approach, I'll remove all
the builtin code from extract_range_builtin, with the exception of the
UBSAN stuff (renaming it to extract_range_ubsan_builtin).
Since the builtin code has proven fickle across architectures, I've
tested this with {-m32,-m64,-fsanitize=signed-integer-overflow} on
x86, ppc64le, and aarch64. I think this should be enough. If it
isn't, we can revert the patch, and leave the duplicate code until
the next release cycle when hopefully vr_values, evrp, and friends
will all be overhauled.
gcc/ChangeLog:
PR tree-optimization/97505
* vr-values.c (vr_values::extract_range_basic): Enable
trap again for everything except UBSAN builtins.
Currently push_template_decl (mostly) decides whether to add a
template header to an entity by seeing if it has DECL_LANG_SPECIFIC.
That might have been a useful predicate at one time, but basing
semantic implications on how we've decided to represent decls is bound
to be brittle. And indeed it is, as more decls grow a use for
lang-specific. In particular I discovered that function-scope
VAR_DECLs couild grow lang-specific, and thereby get a template
header. There's no need for that, and it breaks an invariant modules
was expected.
This patch changes that, and bases the descision on the properties of
the decl. In particular the only function-scope decl that gets a
template header is an implicit-typedef.
I also cleaned up the behaviour of it building a template-info only to
ignore it.
gcc/cp/
* pt.c (push_template_decl): Do not give function-scope entities
other than implicit typedefs a template header. Do not readd
template info to a redeclared template.
libstdc++-v3/ChangeLog:
* testsuite/25_algorithms/lexicographical_compare/constrained.cc:
(test03): Fix initializing the vector vy with the array y of size 4.
For the following test case (compiled with load/store lanes
disabled locally):
void
f (uint32_t *restrict x, uint8_t *restrict y, int n)
{
for (int i = 0; i < n; ++i)
{
x[i * 2] = x[i * 2] + y[i * 2];
x[i * 2 + 1] = x[i * 2 + 1] + y[i * 2];
}
}
we have a redundant no-op permute on the x[] load node:
node 0x4472350 (max_nunits=8, refcnt=2)
stmt 0 _5 = *_4;
stmt 1 _13 = *_12;
load permutation { 0 1 }
Then, when costing it, we pick a cost of 1, even though we need 4 copies
of the x[] load to match a single y[] load:
==> examining statement: _5 = *_4;
Vectorizing an unaligned access.
vect_model_load_cost: unaligned supported by hardware.
vect_model_load_cost: inside_cost = 1, prologue_cost = 0 .
The problem is that the code only considers the permutation for
the first scalar iteration, rather than for all VF iterations.
This patch tries to fix that by making vect_transform_slp_perm_load
calculate the value instead.
gcc/
* tree-vectorizer.h (vect_transform_slp_perm_load): Take an
optional extra parameter.
* tree-vect-slp.c (vect_transform_slp_perm_load): Calculate
the number of loads as well as the number of permutes, taking
the counting loop from...
* tree-vect-stmts.c (vect_model_load_cost): ...here. Use the
value computed by vect_transform_slp_perm_load for ncopies.
