A small tweak to the implementation of __includes, which in my
application saves 20% of the running time. I noticed it because using
range-v3 was giving unexpected performance gains.
Some of the gain comes from pulling the 2 calls ++__first1 out of the
condition so there is just one call. And most of the gain comes from
replacing the resulting
if (__comp(__first1, __first2))
;
else
++__first2;
with
if (!__comp(__first1, __first2))
++__first2;
I was very surprised that the code ended up being so different for such
a change, and I still don't really understand where the extra time is
going...
Anyway, while I blame the compiler for not generating very good code
with the current implementation, I believe the change can be seen as a
simplification.
libstdc++-v3/ChangeLog:
* include/bits/stl_algo.h (__includes): Simplify the code.
The attached patch changes the code generated for
std::optional<std::array<int,1024>>f(){return{};}
from
movq $0, (%rdi)
movq %rdi, %r8
leaq 8(%rdi), %rdi
xorl %eax, %eax
movq $0, 4084(%rdi)
movq %r8, %rcx
andq $-8, %rdi
subq %rdi, %rcx
addl $4100, %ecx
shrl $3, %ecx
rep stosq
movq %r8, %rax
or with different tuning
subq $8, %rsp
movl $4100, %edx
xorl %esi, %esi
call memset
addq $8, %rsp
to the much shorter
movb $0, 4096(%rdi)
movq %rdi, %rax
i.e. the same as the nullopt constructor.
The constructor was already non-trivial, so we don't lose that. It passes the
testsuite without regression, but there is no new testcase to verify the
better codegen.
libstdc++-v3/ChangeLog:
* include/std/optional (optional()): Explicitly define it.
The std::__uninitialized_default_n algorithm used by std::vector creates
an initial object as a local variable then copies that into the
destination range. If the object is too large for the stack this
crashes. We should create the first object directly into the
destination and then copy it from there.
This doesn't fix the bug for C++98, because in that case the initial
value is created as a default argument of the vector constructor i.e. in
the user's code, not inside libstdc++. We can't prevent that.
PR libstdc++/94540
* include/bits/stl_uninitialized.h (__uninitialized_default_1<true>):
Construct the first value at *__first instead of on the stack.
(__uninitialized_default_n_1<true>): Likewise.
Improve comments on several of the non-standard algorithms.
* testsuite/20_util/specialized_algorithms/uninitialized_default/94540.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_default_n/94540.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_value_construct/94540.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_value_construct_n/94540.cc:
New test.
* testsuite/23_containers/vector/cons/94540.cc: New test.
In my recent r11-1460 commit the tests had been "improved" before
commit, and no longer exercised the code paths changed by the patch.
This restores what I originally tested, so that the tests fail before
the r11-1460 change and pass after it.
* testsuite/20_util/specialized_algorithms/uninitialized_default_n/sizes.cc:
Replace Value type with int so trivial code path is used.
* testsuite/20_util/specialized_algorithms/uninitialized_value_construct_n/sizes.cc:
Likewise.
The std::uninitialized_fill_n algorithm uses sd::fill_n for trivial
types, but that algorithm has a stronger requirement that the Size
parameter is convertible to an integral type. As the new test shows,
there are types which are valid for std::uninitialized_fill_n but which
produce a different result when converted to an integral type, or cannot
be converted at all. Only use the std::fill_n optimization when the Size
type is already an integral type.
The std::__uninitialized_default_n extension has the same problem, and
so does C++17's std::uninitialized_value_construct_n which uses it.
* include/bits/stl_uninitialized.h (uninitialized_fill_n): Only
use std::fill_n when the size is an integral type.
(__uninitialized_default_n): Likewise.
* testsuite/20_util/specialized_algorithms/uninitialized_default_n/sizes.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_fill_n/sizes.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_value_construct_n/sizes.cc:
New test.
This improves the previous fix for PR 95282, and extends it to also
apply to the exchange function (which has a similar problem and would
become ill-formed with my proposed fix for PR 95378).
PR libstdc++/95282
* include/bits/atomic_base.h (__atomic_impl::load): Use the _Val
alias instead of deducing _Tp as an unqualified type.
(__atomic_impl::exchange): Use the _Val alias to remove volatile
from the reinterpret_cast result type.
C++20 adds some new preconditions to std::atomic, which weren't
previously checked by our implementation.
* include/std/atomic (atomic): Add static assertions.
* testsuite/29_atomics/atomic/requirements/types_neg.cc: New test.
This patch generalizes our existing functionality for deferring access
checking of typedefs when parsing a function or class template to now
defer all kinds of access checks until template instantiation time,
including member function and member object accesses.
Since all access checks eventually go through enforce_access, the main
component of this patch is new handling inside enforce_access to defer
the current access check if we're inside a template. The bulk of the
rest of the patch consists of removing now-unneeded code pertaining to
suppressing access checks inside templates or pertaining to
typedef-specific access handling. Renamings and other changes with no
functional impact have been split off into the followup patch.
gcc/cp/ChangeLog:
PR c++/41437
PR c++/47346
* call.c (enforce_access): Move to semantics.c.
* cp-tree.h (enforce_access): Delete.
(get_types_needing_access_check): Delete.
(add_typedef_to_current_template_for_access_check): Delete.
* decl.c (make_typename_type): Adjust accordingly. Use
check_accessibility_of_qualified_id instead of directly using
perform_or_defer_access_check.
* parser.c (cp_parser_template_declaration_after_parameters):
Don't push a dk_no_check access state when parsing a template.
* pt.c (get_types_needing_access_check): Delete.
(append_type_to_template_for_access_check_1): Delete.
(perform_typedefs_access_check): Adjust. If type_decl is a
FIELD_DECL, also check its DECL_CONTEXT for dependence. Use
tsubst_copy instead of tsubst to substitute into type_decl so
that we substitute into the DECL_CONTEXT of a FIELD_DECL.
(append_type_to_template_for_access_check): Delete.
* search.c (accessible_p): Remove the processing_template_decl
early exit.
* semantics.c (enforce_access): Moved from call.c. If we're
parsing a template and the access check failed, add the check to
TI_TYPEDEFS_NEEDING_ACCESS_CHECKING.
(perform_or_defer_access_check): Adjust comment.
(add_typedef_to_current_template_for_access_check): Delete.
(check_accessibility_of_qualified_id): Adjust accordingly.
Exit early if the scope is dependent.
gcc/testsuite/ChangeLog:
PR c++/41437
PR c++/47346
* g++.dg/cpp2a/concepts-using2.C: Adjust.
* g++.dg/lto/20081219_1.C: Adjust.
* g++.dg/lto/20091002-1_0.C: Adjust.
* g++.dg/lto/pr65475c_0.C: Adjust.
* g++.dg/opt/dump1.C: Adjust.
* g++.dg/other/pr53574.C: Adjust.
* g++.dg/template/access30.C: New test.
* g++.dg/template/access31.C: New test.
* g++.dg/wrappers/wrapper-around-type-pack-expansion.C: Adjust.
libstdc++-v3/ChangeLog:
PR libstdc++/94003
* testsuite/20_util/is_constructible/94003.cc: New test.
Although not required by SD-6 or the C++20 draft, we define the macro
__cpp_lib_constexpr_char_traits to indicate support for P0432R1. This
updates the value in C++20 mode for the P1032R1 changes to char_traits.
* include/bits/char_traits.h (__cpp_lib_constexpr_char_traits):
Update value for C++20.
* include/std/version (__cpp_lib_constexpr_char_traits): Likewise.
* testsuite/21_strings/char_traits/requirements/constexpr_functions_c++17.cc:
Update expected value.
* testsuite/21_strings/char_traits/requirements/constexpr_functions_c++20.cc:
Likewise.
Upon constexpr evaluation, char_traits move uses copy_backward, but its
last argument should be to the range end rather than its beginning.
2020-06-12 Paul Keir <paul.keir@uws.ac.uk>
* include/bits/char_traits.h (char_traits::move): constexpr move with
overlap was using copy_backward incorrectly.
* testsuite/21_strings/char_traits/requirements/constexpr_functions_c++20.cc:
New test.
The tests for clear() and test_and_set() didn't cover all cases.
* testsuite/29_atomics/atomic_flag/clear/1.cc: Also test clear()
when the value is currently set.
* testsuite/29_atomics/atomic_flag/test_and_set/explicit.cc:
Actually check the return value.
* testsuite/29_atomics/atomic_flag/test_and_set/implicit.cc:
Likewise.
Also fix the tests so they run without an explicit -std=gnu++2a in the
RUNTESTFLAGS, and test the new function on const-qualified objects.
* include/bits/atomic_base.h (atomic_flag::test): Add missing
const qualifiers.
* testsuite/29_atomics/atomic_flag/test/explicit.cc: Add
dg-options and verify results of test function.
* testsuite/29_atomics/atomic_flag/test/implicit.cc: Likewise.
The current code assumes that if the next character in the stream is
equal to the delimiter then we stopped because we saw that delimiter,
and so discards it. But in the testcase for the PR we stop because we
reached the maximum number of characters, and it's coincidence that the
next character equals the delimiter. We should not discard the next
character in that case.
The fix is to check that we haven't discarded __n characters already,
instead of checking whether the next character equals __delim. Because
we've already checked for EOF, if we haven't discarded __n yet then we
know we stopped because we saw the delimiter. On the other hand, if the
next character is the delimiter we don't know if that's why we stopped.
PR libstdc++/94749
* include/bits/istream.tcc (basic_istream::ignore(streamsize, CharT)):
Only discard an extra character if we didn't already reach the
maximum number.
* src/c++98/istream.cc (istream::ignore(streamsiz, char))
(wistream::ignore(streamsize, wchar_t)): Likewise.
* testsuite/27_io/basic_istream/ignore/char/94749.cc: New test.
* testsuite/27_io/basic_istream/ignore/wchar_t/94749.cc: New test.
ranges::copy and a number of other ranges algorithms have unwrapping
optimizations for iterators of type __normal_iterator, move_iterator and
reverse_iterator. But in the checks that guard these optimizations we
currently only test that the iterator of the iterator/sentinel pair has
the appropriate type before proceeding with the corresponding
optimization, and do not also test the sentinel type.
This breaks the testcase in this PR because this testcase constructs via
range adaptors a range whose begin() is a __normal_iterator and whose
end() is a custom sentinel type, and then performs ranges::copy on it.
From there we bogusly perform the __normal_iterator unwrapping
optimization on this iterator/sentinel pair, which immediately leads to
a constraint failure since the custom sentinel type does not model
sentinel_for<int*>.
This patch fixes this issue by refining each of the problematic checks
to also test that the iterator and sentinel types are the same before
applying the corresponding unwrapping optimization. Along the way, some
code simplifications are made.
libstdc++-v3/ChangeLog:
PR libstdc++/95578
* include/bits/ranges_algo.h (__lexicographical_compare_fn):
Also check that the iterator and sentinel have the same type before
applying the unwrapping optimization for __normal_iterator.
Split the check into two, one for the first iterator/sentinel
pair and another for second iterator/sentinel pair. Remove uses
of __niter_base, and remove uses of std::move on a
__normal_iterator.
* include/bits/ranges_algobase.h (__equal_fn): Likewise.
(__copy_or_move): Likewise. Perform similar adjustments for
the reverse_iterator and move_iterator optimizations. Inline
the checks into the if-constexprs, and use using-declarations to
make them less visually noisy. Remove uses of __niter_wrap.
(__copy_or_move_backward): Likewise.
* testsuite/25_algorithms/copy/95578.cc: New test.
* testsuite/25_algorithms/copy_backward/95578.cc: New test.
* testsuite/25_algorithms/equal/95578.cc: New test.
* testsuite/25_algorithms/lexicographical_compare/95578.cc: New test.
* testsuite/25_algorithms/move/95578.cc: New test.
* testsuite/25_algorithms/move_backward/95578.cc: New test.
Make the memcmp optimization work for std::deque iterators and safe
iterators.
Co-authored-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
2020-06-08 François Dumont <fdumont@gcc.gnu.org>
Jonathan Wakely <jwakely@redhat.com>
* include/bits/deque.tcc (__lex_cmp_dit): New.
(__lexicographical_compare_aux1): Define overloads for deque
iterators.
* include/bits/stl_algobase.h (__lexicographical_compare::__3way):
New static member function.
(__lexicographical_compare<true>::__3way): Likewise.
(__lexicographical_compare<true>::__lc): Use __3way.
(__lexicographical_compare_aux): Rename to
__lexicographical_compare_aux1 and declare overloads for deque
iterators.
(__lexicographical_compare_aux): Define new forwarding function
that calls __lexicographical_compare_aux1 and declare new overloads
for safe iterators.
(lexicographical_compare): Do not use __niter_base on
parameters.
* include/debug/safe_iterator.tcc
(__lexicographical_compare_aux): Define overloads for safe
iterators.
* testsuite/25_algorithms/lexicographical_compare/1.cc: Add
checks with random access iterators.
* testsuite/25_algorithms/lexicographical_compare/deque_iterators/1.cc:
New test.
As clarified by LWG 3265, std::move_iterator is supposed to have an
assignment operator that converts from a different specialization of
std::move_iterator, which performs an assignment. That has always been
missing from libstdc++, so assigning a different type actually performs
a converting construction, then an assignment. This is non-conforming
for the (fairly contrived) case where the converting assignment is
well-formed but the converting construction is not.
* include/bits/stl_iterator.h (move_iterator::operator=): Define.
* testsuite/24_iterators/move_iterator/dr3265.cc: New test.
The standard requires that std::bad_optional_access' default
constructor has a non-throwing exception specification.
* include/std/optional (bad_optional_access): Define default
constructor and destructor as defaulted.
* testsuite/20_util/optional/bad_access.cc: New test.
These started failing with the previous commit, because I forgot to add
the tests after adjusting them.
* testsuite/20_util/default_delete/48631_neg.cc: Adjust dg-error
line number.
* testsuite/20_util/default_delete/void_neg.cc: Likewise.
With PR c++/92078 and PR c++/92103 both fixed, nested class templates
can now be constrained. That means a number of namespace-scope helpers
can be moved to the class scope, so they're only visible where they're
needed.
* include/bits/iterator_concepts.h (__detail::__ptr, __detail::__ref)
(__detail::__cat, __detail::__diff): Move to class scope in the
relevant __iterator_traits specializations.
(__iterator_traits<>): Use nested class templates instead of ones from
namespace __detail.
* include/bits/stl_iterator.h (__detail::__common_iter_ptr): Move to
class scope in iterator_traits<common_iterator<I, S>>.
(iterator_traits<common_iterator<I, S>>): Use nested class template
instead of __detail::__common_iter_ptr.
Since it was added in C++11, std::copy_n and std::ranges::copy_n should
do nothing given a negative size, but for random access iterators we add
the size to the iterator, possibly resulting in undefined behaviour.
Also, C++20 clarified that std::copy_n requires the Size type to be
convertible to an integral type. We previously assumed that it could be
directly used in arithmetic expressions, without conversion to an
integral type.
This also fixes a bug in the random_access_iterator_wrapper helper adds
some convenience aliases for using the iterator wrappers.
libstdc++-v3/ChangeLog:
* include/bits/ranges_algobase.h (__copy_n_fn): Only call
ranges::copy for positive values.
* include/bits/stl_algo.h (copy_n): Convert Size argument to an
integral type and only call __copy_n for positive values.
* testsuite/util/testsuite_iterators.h
(random_access_iterator_wrapper::operator+=): Fix range check for
negative values.
(output_container, input_container, forward_container)
(bidirectional_container, random_access_container): New alias
templates.
* testsuite/25_algorithms/copy_n/5.cc: New test.
When given a type which can convert to any container-like type, the
C(const C&) copy constructor and C(const C::_Base&) converting
constructor are ambiguous. This change replaces the converting
constructor's parameter with a reference_wrapper-like type so that
calling that constructor requires an additional user-defined conversion.
This gives it a lower rank than the copy constructor, avoiding the
ambiguity.
While testing this change I discovered that __gnu_debug::forward_list
doesn't have a convering constructor from the std::forward_list base, so
this adds it.
We should probably consider whether the converting constructors should
be 'explicit' but I'm not changing that now.
libstdc++-v3/ChangeLog:
PR libstdc++/90102
* include/debug/deque (deque(const _Base&)): Replace parameter
with a struct that wraps a const _Base&.
* include/debug/forward_list (forward_list(_Base_ref)): New
constructor.
* include/debug/list (list(const _Base&)): Replace parameter
with a struct that wraps a const _Base&.
* include/debug/map.h (map(const _Base&)): Likewise.
* include/debug/multimap.h (multimap(const _Base&)): Likewise.
* include/debug/multiset.h (multiset(const _Base&)): Likewise.
* include/debug/set.h (set(const _Base&)): Likewise.
* include/debug/unordered_map (unordered_map(const _Base&))
(unordered_multimap(const _Base&)): Likewise.
* include/debug/unordered_set (unordered_set(const _Base&))
(unordered_multiset(const _Base&)): Likewise.
* testsuite/23_containers/vector/cons/destructible_debug_neg.cc:
Adjust dg-error line number.
* include/debug/vector (vector(const _Base&)): Likewise.
* testsuite/23_containers/deque/debug/90102.cc: New test.
* testsuite/23_containers/forward_list/debug/90102.cc: New test.
* testsuite/23_containers/list/debug/90102.cc: New test.
* testsuite/23_containers/map/debug/90102.cc: New test.
* testsuite/23_containers/multimap/debug/90102.cc: New test.
* testsuite/23_containers/multiset/debug/90102.cc: New test.
* testsuite/23_containers/set/debug/90102.cc: New test.
* testsuite/23_containers/unordered_map/debug/90102.cc: New test.
* testsuite/23_containers/unordered_multimap/debug/90102.cc: New test.
* testsuite/23_containers/unordered_multiset/debug/90102.cc: New test.
* testsuite/23_containers/unordered_set/debug/90102.cc: New test.
* testsuite/23_containers/vector/debug/90102.cc: New test.
When I refactored filesystem::path string conversions in
r11-587-584d52b088f9fcf78704b504c3f1f07e17c1cded I failed to update the
mingw-specific code in filesystem::u8path, causing a bootstrap failure.
This fixes it, and further refactors the mingw-specific code along the
same lines as the previous commit. All conversions from UTF-8 strings to
wide strings now use the same helper function, __wstr_from_utf8.
PR libstdc++/95392
* include/bits/fs_path.h (path::_S_to_string): Move to
namespace-scope and rename to ...
(__detail::__string_from_range): ... this.
[WINDOWS] (__detail::__wstr_from_utf8): New function template to
convert a char sequence containing UTF-8 to wstring.
(path::_S_convert(Iter, Iter)): Adjust call to _S_to_string.
(path::_S_convert_loc(Iter, Iter, const locale&)): Likewise.
(u8path(InputIterator, InputIterator)) [WINDOWS]: Use
__string_from_range to obtain a contiguous range and
__wstr_from_utf8 to obtain a wide string.
(u8path(const Source&)) [WINDOWS]: Use __effective_range to
obtain a contiguous range and __wstr_from_utf8 to obtain a wide
string.
(path::_S_convert(const _EcharT*, const _EcharT)) [WINDOWS]:
Use __wstr_from_utf8.
I noticed recently that our input_iterator_wrapper utility for writing
tests has the following post-increment operator:
void
operator++(int)
{
++*this;
}
That fails to meet the Cpp17InputIterator requirement that *r++ is
valid. This change makes it return a non-void proxy type that can be
deferenced to produce another proxy, which is convertible to the
value_type. The second proxy converts to const T& to ensure it can't be
written to.
* testsuite/util/testsuite_iterators.h:
(input_iterator_wrapper::operator++(int)): Return proxy object.
The <xref> element creates the link text automatically from the link
target, rather than using the text node child of the element. This can
be changed by using an endterm attribute, but it's simpler to just use
the <link> element instead.
* doc/xml/manual/containers.xml: Replace <xref> with <link>.
* doc/xml/manual/evolution.xml: Likewise.
* doc/html/manual/api.html: Regenerate.
* doc/html/manual/containers.html: Regenerate.
There is a stray change (introducing a bogus line at the top) that
came via 2babbb6760c43bcd803a5e168ef5ecb0be8a5121; remove that again.
* doc/xml/manual/policy_data_structures_biblio.xml: Remove
stray change.
Use the GLIBCXX_CHECK_MATH_DECL macro to check for the full list of
vxworks math decls.
libstdc++-v3/ChangeLog
* crossconfig.m4 (<*-vxworks>): Check for more math decls.
* configure: Rebuild.
Those methods are making a double lookup in case of insertion, they can
perform only one.
PR libstdc++/95079
* include/bits/hashtable_policy.h (_Insert_base<>::try_emplace): New.
* include/bits/unordered_map.h (unordered_map<>::try_emplace): Adapt.
(unordered_map<>::insert_or_assign): Adapt.
