[Bug c++/109753] New: pragma GCC target stops std::vector from compiling

[Bug c++/109753] New: pragma GCC target stops std::vector from compiling

[magnus.hegdahl at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

            Bug ID: 109753
           Summary: pragma GCC target stops std::vector from compiling
           Product: gcc
           Version: 13.1.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: magnus.hegdahl at gmail dot com
  Target Milestone: ---

Tested on g++ (GCC) 13.1.1 20230429 on x86-64 Linux

The following code doesn't compile without also enabling the target using
compile flags (like -mavx2). It did compile in GCC 12.2.0.

#pragma GCC target("avx2")

#include <vector>

int main() {
  std::vector<int> a;
}

[Bug target/109753] [13/14 Regression] pragma GCC target stops std::vector from compiling

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |13.2

[Bug target/109753] [13/14 Regression] pragma GCC target stops std::vector from compiling

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
reducing ...

[Bug target/109753] [13/14 Regression] pragma GCC target stops std::vector from compiling

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-05-05

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Reduced:
```
#pragma GCC target("avx2")
struct aa {
    __attribute__((__always_inline__)) aa() noexcept {}
};
aa _M_impl;
```

[Bug target/109753] [13/14 Regression] pragma GCC target stops std::vector from compiling

[mpolacek at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Marek Polacek <mpolacek at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mpolacek at gcc dot gnu.org

--- Comment #3 from Marek Polacek <mpolacek at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #2)
> Reduced:
> ```
> #pragma GCC target("avx2")
> struct aa {
>     __attribute__((__always_inline__)) aa() noexcept {}
> };
> aa _M_impl;
> ```

This started to give an error with

commit bef8491a658de9e8920acaeff6cb76ef4e946e2c
Author: Sriraman Tallam <tmsriram@google.com>
Date:   Tue Jun 18 22:45:03 2013 +0000

    Emit errors when always_inline functions cannot be inlined in -O0 mode.

            * tree-inline.c (expand_call_inline): Allow the error to be flagged
            in early inline pass.
            * ipa-inline.c (inline_always_inline_functions): Pretend
always_inline
            functions are inlined during failures to flag an error.
            * gcc.target/i386/inline_error.c: New test.
            * gcc.c-torture/compile/pr44043.c: Fix test to expect an error.
            * gcc.c-torture/compile/pr43791.c: Fix test to expect an error.

    From-SVN: r200179

I think the original problem must have been triggered by some libstdc++ change.

[Bug target/109753] [13/14 Regression] pragma GCC target stops std::vector from compiling

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Marek Polacek from comment #3)
> I think the original problem must have been triggered by some libstdc++
> change.

Yes r13-3816-g9d549401ae8ab3 .

[Bug target/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I suspect the generated constructor for the clones is not getting the target
options on it like it should be ...

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |c++
                 CC|                            |rguenth at gcc dot gnu.org
           Priority|P3                          |P2

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
That would be a C++ frontend issue then, no?  Maybe it's also only some of the
aliases?  And the actual issue would be latent then.

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[xry111 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Xi Ruoyao <xry111 at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |gcc.gnu.org at aryanc403 dot com

--- Comment #7 from Xi Ruoyao <xry111 at gcc dot gnu.org> ---
*** Bug 110675 has been marked as a duplicate of this bug. ***

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[xry111 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Xi Ruoyao <xry111 at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |xry111 at gcc dot gnu.org

--- Comment #8 from Xi Ruoyao <xry111 at gcc dot gnu.org> ---
In local-fnsummary2:

__attribute__((always_inline, target ("avx2")))
void aa::aa (struct aa * const this)
{
  <bb 2> [local count: 1073741824]:
  return;

}

this seems correct.  But:

void __static_initialization_and_destruction_0 ()
{
  <bb 2> :
  aa::aa (&_M_impl);
  return;

}

Note that __static_initialization_and_destruction_0 is not attributed with
avx2.

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Xi Ruoyao from comment #8)
> In local-fnsummary2:
> 
> __attribute__((always_inline, target ("avx2")))
> void aa::aa (struct aa * const this)
> {
>   <bb 2> [local count: 1073741824]:
>   return;
> 
> }
> 
> this seems correct.  But:
> 
> void __static_initialization_and_destruction_0 ()
> {
>   <bb 2> :
>   aa::aa (&_M_impl);
>   return;
> 
> }
> 
> Note that __static_initialization_and_destruction_0 is not attributed with
> avx2.

I think that's correct.  Maybe we need multiple CTOR/DTOR functions
in such case.

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|13.2                        |13.3

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 13.2 is being released, retargeting bugs to GCC 13.3.

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Martin Jambor <jamborm at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|needs-bisection             |

--- Comment #11 from Martin Jambor <jamborm at gcc dot gnu.org> ---
It seems there is nothing to bisect any more, please re-add the keyword if I am
wrong.

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #12 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
I think this is a problem with two meanings of always_inline. One is "it must
be inlined or otherwise we will not be able to generate code" other is
"disregard inline limits".

I guess practical solution here would be to ingore always inline for functions
called from static construction wrappers (since they only optimize around array
of function pointers). Question is how to communicate this down from FE to
ipa-inline...

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think the issue might be that whoever is creating
__static_initialization_and_destruction_0 fails to honor the active
target pragma.  Which means back to my suggestion to have multiple ones
when different target options are on the individual CTORs and any of them
have always-inline (with always-inline we can't rely on an out-of-line copy
to exist).

Yes, for libstdc++ purposes which seems to get more and more "always-inline"
it would be good to have a different attribute that would only cause to
disregard inline limits and not have "always-inline" semantics.

[[inline_without_limits]] or [[inline_no_limits]]

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[hubicka at ucw dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #14 from Jan Hubicka <hubicka at ucw dot cz> ---
> I think the issue might be that whoever is creating
> __static_initialization_and_destruction_0 fails to honor the active
> target pragma.  Which means back to my suggestion to have multiple ones
> when different target options are on the individual CTORs and any of them
> have always-inline (with always-inline we can't rely on an out-of-line copy
> to exist).

If I remember right, __static_initialization_and_destruction_0 call all
static constructors of priority 0. So it has really no active pragma
since it may change across translation unit.

We also have similar code in IPA where we collect constructors
across whole program.  Motivation was to get them inlined and obtain
better code locality. Back then Firefox had iostram constructor in every
object file and those ctors made whole text segment to be loaded on
demand during startup.

Probably correct solution would be to group construtor into groups by
compatible optimization/target pgramas in the inliner's definition.
A quick hack would be to generate wrapper calls which will "eat" the
always inline...

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[jason at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

Jason Merrill <jason at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jason at gcc dot gnu.org
                 CC|                            |jason at gcc dot gnu.org

[Bug c++/109753] [13/14 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[jason at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #15 from Jason Merrill <jason at gcc dot gnu.org> ---
Created attachment 57706
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=57706&action=edit
one approach

I tried just making implicit functions respect #pragma target, but that
regresses pr105306 due to seeming internal confusion over whether -Ofast or
#pragma optimize apply to the implicit ~C.  I haven't tracked that down yet.

[Bug c++/109753] [13/14/15 Regression] pragma GCC target causes std::vector not to compile (always_inline on constructor)

[jason at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109753

--- Comment #16 from Jason Merrill <jason at gcc dot gnu.org> ---
Patch posted: https://gcc.gnu.org/pipermail/gcc-patches/2024-May/650439.html