https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101786

            Bug ID: 101786
           Summary: P1143R2 constinit implementation is incomplete
                    (joining with thread_local)
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jpegqs at gmail dot com
                CC: mpolacek at gcc dot gnu.org
  Target Milestone: ---

The paper says:

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1143r2.html

> constinit can also be useful to compilers for non-initializing declarations
> of thread_local variables:
> 
> extern thread_local constinit x;
> int f() { return x; }
> 
> Without constinit, runtime code must be executed to perform a check of a
> guard variable and conditionally initialize x each time it is used. (Other
> techniques exist, but this approach is common.) If the variable is known to
> have constant initialization, this can be avoided.

Let's fix the missing type for x and try:

extern thread_local constinit int x;
int f() { return x; }

In case of compilation, GCC does not remove the TLS wrapper function as it
should according to this paper:

_ZTW1x:
        push    rbp
        mov     rbp, rsp
        mov     eax, OFFSET FLAT:_ZTH1x
        test    rax, rax
        je      .L2
        call    _ZTH1x
.L2:
        mov     rdx, QWORD PTR fs:0
        mov     rax, QWORD PTR x@gottpoff[rip]
        add     rax, rdx
        pop     rbp
        ret
_Z1fv:
        push    rbp
        mov     rbp, rsp
        call    _ZTW1x
        mov     eax, DWORD PTR [rax]
        pop     rbp
        ret

The code it should produce should look like this: 

_Z1fv:
        push    rbp
        mov     rbp, rsp
        mov     rax, QWORD PTR x@gottpoff[rip]
        mov     eax, DWORD PTR fs:[rax]
        pop     rbp
        ret

What I can get now is only by replacing "thread_local constinit" with
"__thread".

Clang implements this feature.

Reply via email to