On Fri, 29 Sept 2023 at 17:29, Nathaniel Shead
<nathanielosh...@gmail.com> wrote:
>
> On Fri, Sep 29, 2023 at 04:06:33PM +0100, Jonathan Wakely wrote:
> > On Fri, 29 Sept 2023 at 10:32, Jonathan Wakely <jwak...@redhat.com> wrote:
> > > > Thanks for the comments, here's an updated version of the patch.
> > > > Bootstrapped and regtested on x86_64-pc-linux-gnu.
> > >
> > > Great, I'll get this committed today - thanks!
> >
> > That's done now.
> >
>
> Thanks!
>
> > > >
> > > > I'll note that there are some existing calls to `_M_use_local_data()`
> > > > already used only for their side effects without a cast to void, e.g.
> > > >
> > > > /**
> > > > * @brief Default constructor creates an empty string.
> > > > */
> > > > _GLIBCXX20_CONSTEXPR
> > > > basic_string()
> > > >
> > > > _GLIBCXX_NOEXCEPT_IF(is_nothrow_default_constructible<_Alloc>::value)
> > > > : _M_dataplus(_M_local_data())
> > > > {
> > > > _M_use_local_data();
> > > > _M_set_length(0);
> > > > }
> > > >
> > > > I haven't updated these, but should this be changed for consistency?
> > >
> > > Yes, good idea. I can do that.
> >
> > I started to do that, and decided it made more sense to split out the
> > constexpr loop from _M_use_local_data() into a separate function,
> > _M_init_local_buf(). Then we can use that for all the places where we
> > don't care about the return value. That avoids the overhead of using
> > pointer_traits::pointer_to when we don't need the return value (which
> > is admittedly only going to be an issue for -O0 code, but I think it's
> > cleaner this way anyway).
> >
> > Please see the attached patch and let me know what you think.
>
> I agree, and it also looks clearer to me what is happening.
Good, I'll make this change next week then.
>
> >
> > > Thanks again for fixing these. I think this might fix some bug reports
> > > about clang rejecting our std::string in constant expressions, so I'll
> > > check those.
> >
> > Your patch fixes https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110900
> > (so we should backport it to gcc-13 and gcc-12 too).
>
> > commit 2668979d3206ff6c039ac0165aae29377a15666c
> > Author: Jonathan Wakely <jwak...@redhat.com>
> > Date: Fri Sep 29 12:12:22 2023
> >
> > libstdc++: Split std::basic_string::_M_use_local_data into two functions
> >
> > This splits out the activate-the-union-member-for-constexpr logic from
> > _M_use_local_data, so that it can be used separately in cases that don't
> > need to use std::pointer_traits<pointer>::pointer_to to obtain the
> > return value.
> >
> > This leaves only three uses of _M_use_local_data() which are all the
> > same form:
> >
> > __s._M_data(_M_use_local_data());
> > __s._M_set_length(0);
> >
> > We could remove _M_use_local_data() and change those three places to use
> > a new _M_reset() function that does:
> >
> > _M_init_local_buf();
> > _M_data(_M_local_data());
> > _M_set_length(0);
> >
> > This is left for a future change.
> >
> > libstdc++-v3/ChangeLog:
> >
> > * include/bits/basic_string.h (_M_init_local_buf()): New
> > function.
> > (_M_use_local_data()): Use _M_init_local_buf.
> > (basic_string(), basic_string(const Alloc&))
> > (basic_string(basic_string&&))
> > (basic_string(basic_string&&, const Alloc&)): Use
> > _M_init_local_buf instead of _M_use_local_data().
> > * include/bits/basic_string.tcc (swap(basic_string&))
> > (_M_construct(InIter, InIter, forward_iterator_tag))
> > (_M_construct(size_type, CharT), reserve()): Likewise.
> > (_M_construct(InIter, InIter, input_iterator_tag)): Remove call
> > to _M_use_local_data() and initialize the local buffer directly.
> >
> > diff --git a/libstdc++-v3/include/bits/basic_string.h
> > b/libstdc++-v3/include/bits/basic_string.h
> > index 4f94cd967cf..18a19b8dcbc 100644
> > --- a/libstdc++-v3/include/bits/basic_string.h
> > +++ b/libstdc++-v3/include/bits/basic_string.h
> > @@ -353,13 +353,23 @@ _GLIBCXX_BEGIN_NAMESPACE_CXX11
> > // Ensure that _M_local_buf is the active member of the union.
> > __attribute__((__always_inline__))
> > _GLIBCXX14_CONSTEXPR
> > - pointer
> > - _M_use_local_data() _GLIBCXX_NOEXCEPT
> > + void
> > + _M_init_local_buf() _GLIBCXX_NOEXCEPT
> > {
> > #if __cpp_lib_is_constant_evaluated
> > if (std::is_constant_evaluated())
> > for (size_type __i = 0; __i <= _S_local_capacity; ++__i)
> > _M_local_buf[__i] = _CharT();
> > +#endif
> > + }
> > +
> > + __attribute__((__always_inline__))
> > + _GLIBCXX14_CONSTEXPR
> > + pointer
> > + _M_use_local_data() _GLIBCXX_NOEXCEPT
> > + {
> > +#if __glibcxx_is_constant_evaluated
> > + _M_init_local_buf();
> > #endif
> > return _M_local_data();
> > }
>
> What's the difference between __cpp_lib_is_constant_evaluated and
> __glibcxx_is_constant_evaluated? Should these lines be using the same
> macro here?
Ah, yeah, they could be the same.
The difference is that (for most feature test macros) the
__glibcxx_ftm form is defined after including <bits/version.h>, but
the __cpp_lib_ftm form is only defined only by headers that do:
#define __glibcxx_want_ftm
#include <bits/version.h>
This means we can ensure that the __cpp_lib_ftm form is only defined
by headers where we actually want to define it, e.g. in the headers
that [version.syn] in the standard says should define the macro. So
for our own internal uses, we should generally rely on the
__glibcxx_ftm one. Users should rely on __cpp_lib_ftm after including
the correct header. For example, __glibcxx_atomic_wait is defined
after including <bits/atomic_wait.h> and so is available for our own
uses, but __cpp_lib_atomic_wait is only defined after including
<atomic>, not just <bits/atomic_wait.h>. Or at least, that's the plan
- I have a pending patch to make everything I just said true :-)
Currently we "leak" the __cpp_lib_ftm forms into lots of <bits/xxx.h>
internal headers. That will change next week.
In this specific case, it doesn't matter, because
__cpp_lib_is_constant_evaluated is defined by <type_traits>, and every
header includes that. So it doesn't really matter whether our internal
uses are __cpp_lib_is_constant_evaluated or
__glibcxx_is_constant_evaluated. But it would be good to be
consistent.
>
> > @@ -522,7 +532,7 @@ _GLIBCXX_BEGIN_NAMESPACE_CXX11
> > _GLIBCXX_NOEXCEPT_IF(is_nothrow_default_constructible<_Alloc>::value)
> > : _M_dataplus(_M_local_data())
> > {
> > - _M_use_local_data();
> > + _M_init_local_buf();
> > _M_set_length(0);
> > }
> >
> > @@ -534,7 +544,7 @@ _GLIBCXX_BEGIN_NAMESPACE_CXX11
> > basic_string(const _Alloc& __a) _GLIBCXX_NOEXCEPT
> > : _M_dataplus(_M_local_data(), __a)
> > {
> > - _M_use_local_data();
> > + _M_init_local_buf();
> > _M_set_length(0);
> > }
> >
> > @@ -678,7 +688,7 @@ _GLIBCXX_BEGIN_NAMESPACE_CXX11
> > {
> > if (__str._M_is_local())
> > {
> > - (void)_M_use_local_data();
> > + _M_init_local_buf();
> > traits_type::copy(_M_local_buf, __str._M_local_buf,
> > __str.length() + 1);
> > }
> > @@ -718,7 +728,7 @@ _GLIBCXX_BEGIN_NAMESPACE_CXX11
> > {
> > if (__str._M_is_local())
> > {
> > - (void)_M_use_local_data();
> > + _M_init_local_buf();
> > traits_type::copy(_M_local_buf, __str._M_local_buf,
> > __str.length() + 1);
> > _M_length(__str.length());
> > diff --git a/libstdc++-v3/include/bits/basic_string.tcc
> > b/libstdc++-v3/include/bits/basic_string.tcc
> > index 4bc98f2aea7..052eeb9e846 100644
> > --- a/libstdc++-v3/include/bits/basic_string.tcc
> > +++ b/libstdc++-v3/include/bits/basic_string.tcc
> > @@ -79,7 +79,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> > }
> > else if (__s.length())
> > {
> > - (void)_M_use_local_data();
> > + _M_init_local_buf();
> > traits_type::copy(_M_local_buf, __s._M_local_buf,
> > __s.length() + 1);
> > _M_length(__s.length());
> > @@ -88,7 +88,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> > }
> > else if (length())
> > {
> > - (void)__s._M_use_local_data();
> > + __s._M_init_local_buf();
> > traits_type::copy(__s._M_local_buf, _M_local_buf,
> > length() + 1);
> > __s._M_length(length());
> > @@ -99,7 +99,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> > else
> > {
> > const size_type __tmp_capacity = __s._M_allocated_capacity;
> > - (void)__s._M_use_local_data();
> > + __s._M_init_local_buf();
> > traits_type::copy(__s._M_local_buf, _M_local_buf,
> > length() + 1);
> > _M_data(__s._M_data());
> > @@ -111,7 +111,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> > const size_type __tmp_capacity = _M_allocated_capacity;
> > if (__s._M_is_local())
> > {
> > - (void)_M_use_local_data();
> > + _M_init_local_buf();
> > traits_type::copy(_M_local_buf, __s._M_local_buf,
> > __s.length() + 1);
> > __s._M_data(_M_data());
> > @@ -174,14 +174,18 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> > size_type __len = 0;
> > size_type __capacity = size_type(_S_local_capacity);
> >
> > - pointer __p = _M_use_local_data();
> > -
> > while (__beg != __end && __len < __capacity)
> > {
> > - __p[__len++] = *__beg;
> > + _M_local_buf[__len++] = *__beg;
> > ++__beg;
> > }
> >
> > +#if __glibcxx_is_constant_evaluated
> > + if (std::is_constant_evaluated())
> > + for (size_type __i = __len; __i <= __capacity; ++__i)
> > + _M_local_buf[__i] = _CharT();
> > +#endif
> > +
>
> I wonder if maybe this should still be a call to `_M_init_local_buf()`
> above, where the `_M_use_local_data()` used to be? That way the logic
> stays in one place, and I don't imagine the compile time savings of not
> immediately overwriting the first __len characters would be significant.
Yeah, I went back and forth on that, but you're probably right. I'll
do as you suggested.
