[PATCH] c++, libstdc++, v2: Implement C++26 P3074R7, P3726R0 and CWG3189 - trivial unions [PR119059]

[Tomasz Kaminski]

On Thu, May 21, 2026 at 10:51 AM Jakub Jelinek <jakub@redhat.com> wrote:

> On Thu, May 21, 2026 at 10:31:25AM +0200, Tomasz Kaminski wrote:
> > Even with this patch, I still see a difference between the foo and bar
> > functions.
> > as defined below:
> > #include <memory>
> > #include <cassert>
> > #include <string>
> >
> > template <typename T, size_t N>
> > struct FixedVector {
> >   union { T storage[N]; };
> >   size_t size = 0;
> >
> >   constexpr FixedVector () { std::start_lifetime (storage); }
> >
> >   constexpr ~FixedVector() { std::destroy(storage, storage + size); }
> >
> >   constexpr FixedVector(std::initializer_list<T> il)
> >   {
> >     assert(il.size() < N);
> >     std::start_lifetime (storage);
> >     std::uninitialized_copy(il.begin(), il.end(), storage);
> >     size = il.size();
> >   }
> >
> >   constexpr void push_back(T const& v) { ::new (storage + size) T(v);
> > ++size; }
> > };
> >
> > FixedVector<char, 4> foo()
> > {
> >   return {'a', 'b', 'c'};
> > }
> >
> > FixedVector<char, 4> bar()
> > {
> >   constexpr FixedVector<char, 4> res{'a', 'b', 'c'};
> >   return res;
> > }
> >
> > After compiling the above with -O2 I see:
>
> This boils down to the differences in C with -O2 on:
> struct S { unsigned char a[4]; unsigned long b; };
>
> struct S foo () { static const unsigned char c[3] = { 97, 98, 99 }; struct
> S r; __builtin_memcpy (&r.a, &c, 3); r.b = 3; return r; }
> struct S bar () { struct S r; r.a[0] = 97; r.a[1] = 98; r.a[2] = 99; r.b =
> 3; return r; }
> struct S baz () { static const unsigned char c[4] = { 97, 98, 99, 100 };
> struct S r; __builtin_memcpy (&r.a, &c, 4); r.b = 4; return r; }
> struct S qux () { struct S r; r.a[0] = 97; r.a[1] = 98; r.a[2] = 99;
> r.a[3] = 100; r.b = 4; return r; }
>
> Note, baz and qux emit the same
>         movl    $4, %edx
>         movl    $1684234849, %eax
> while
>         movzbl  .LC0+1(%rip), %edx
>         movzbl  .LC0(%rip), %eax
>         salq    $8, %rdx
>         orq     %rax, %rdx
>         movzbl  .LC0+2(%rip), %eax
>         salq    $16, %rax
>         orq     %rdx, %rax
>         movl    $3, %edx
>         ret
> vs.
>         movl    $3, %edx
>         movl    $6513249, %eax
>         ret
> I guess worth filing a PR, but I'm afraid we have plenty of prior PRs
> about this too (and some attempt to improve it but not really successful;
> it is about the returning of uninitialized bits in registers).
>
That what I assumed with current inplace_vector, that zeroes all elements
of array at compile time. But with the start_lifetime, constexpr
FixedVector<char, 4> res{'a', 'b', 'c'};
It also contains the same out of lifetime element, as return {'a', 'b',
'c'} would.

I think the expecation from
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124121
is that `return FixedVector<char, 4>{'a', 'b', 'c'}` would constant-fold
creation of
fixed vector, and be equivalent to `constexpr FixedVector<char, 4> res{'a',
'b', 'c'}; return res;`.


>
>         Jakub
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://gcc.gnu.org/pipermail/libstdc++/attachments/20260521/7c28abc8/attachment.htm>