[Bug libstdc++/109816] New: [14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[Bug libstdc++/109816] New: [14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

            Bug ID: 109816
           Summary: [14 Regression] r14-321 fix for PR108969 breaks NVPTX
                    offloading due to __asm (".globl
                    _ZSt21ios_base_library_initv")
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libstdc++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: burnus at gcc dot gnu.org
                CC: jakub at gcc dot gnu.org
  Target Milestone: ---

Commit r14-321-g9a41d2cdbcd2af77a3a91a840a3a13f0eb39971b
"libstdc++: Another attempt to ensure g++ 13+ compiled programs enforce gcc
13.2+ libstdc++.so.6 [PR108969]"

breaks some offload compiling with nvptx.

For instance,
https://github.com/TApplencourt/OvO/blob/master/test_src/cpp/hierarchical_parallelism/reduction_add-complex_double/target__teams__parallel_for_simd.cpp

fails to assemble with ptxas (either at compile time or at runtime) with:

ptxas a.xnvptx-none.mkoffload.s, line 8; fatal   : Parsing error near '.globl':
syntax error

The problem is the code:

        // #APP 
        .globl _ZSt21ios_base_library_initv
        // #NO_APP 

Which the commit added to the header file as:

 #if !__has_attribute(__init_priority__)
   static ios_base::Init __ioinit;
+#elif defined(_GLIBCXX_SYMVER_GNU)
+  __extension__ __asm (".globl _ZSt21ios_base_library_initv");
 #endif

 * * *

NOTE: With GCN offloading the program runs, presumably because the LLVM
assembler (mc) does not choke on ".globl".

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Then _GLIBCXX_SYMVER_GNU should not be defined for NVPTX ....

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> Then _GLIBCXX_SYMVER_GNU should not be defined for NVPTX ....

It isn't – but with offloading, we only parse the file once — and then only
later generate for offloading regions code both for the host (→ host fallback)
and for the device.

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
And we emit all toplevel asms into the offloading target code?
Or how does it make into PTX?

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #4 from Tobias Burnus <burnus at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #3)
> And we emit all toplevel asms into the offloading target code?
> Or how does it make into PTX?

It seems as if this is always written (once). Thus, the minimal change would be
the following. We could save some bits by not writing lto_output_toplevel_asms
without '-flto', but I think that is not really needed, given that top-level
asm are rather rare and small.

--- a/gcc/lto-cgraph.cc
+++ b/gcc/lto-cgraph.cc
@@ -1587,3 +1587,5 @@ input_cgraph_1 (struct lto_file_decl_data *file_data,

+#ifndef ACCEL_COMPILER
   lto_input_toplevel_asms (file_data, file_data->order_base);
+#endif

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Tobias Burnus from comment #4)
> (In reply to Jakub Jelinek from comment #3)
> > And we emit all toplevel asms into the offloading target code?
> > Or how does it make into PTX?
> 
> It seems as if this is always written (once). Thus, the minimal change would
> be the following. We could save some bits by not writing
> lto_output_toplevel_asms without '-flto', but I think that is not really
> needed, given that top-level asm are rather rare and small.
> 
> --- a/gcc/lto-cgraph.cc
> +++ b/gcc/lto-cgraph.cc
> @@ -1587,3 +1587,5 @@ input_cgraph_1 (struct lto_file_decl_data *file_data,
>  
> +#ifndef ACCEL_COMPILER
>    lto_input_toplevel_asms (file_data, file_data->order_base);
> +#endif

The above can work only if toplevel asms are in separate section and so
inputting it or not doesn't affect input of other data.
I think it would be better to also not to stream it if lto_stream_offload_p.

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #6 from Tobias Burnus <burnus at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #5)
> The above can work only if toplevel asms are in separate section and so
> inputting it or not doesn't affect input of other data.
> I think it would be better to also not to stream it if lto_stream_offload_p.

The toplevel asms are in a special section: LTO_section_asm — and that section
(enum value) is only used by lto_output_toplevel_asms/lto_input_toplevel_asms.

The output symbols are in symtab->first_asm_symbol(); and all symbols added
there are added through symbol_table::finalize_toplevel_asm(t).


However, if you don't want to stream it out, we could do the following to
prevent it:

diff --git a/gcc/lto-streamer-out.cc b/gcc/lto-streamer-out.cc
index 0bca530..853688b 100644
--- a/gcc/lto-streamer-out.cc
+++ b/gcc/lto-streamer-out.cc
@@ -2525,3 +2525,4 @@ lto_output_toplevel_asms (void)

-  if (!symtab->first_asm_symbol ())
+  /* Skip if there is no LTO but only flag_generate_offload.  */
+  if ((!in_lto_p && !flag_lto) || !symtab->first_asm_symbol ())
     return;
--- a/gcc/lto-cgraph.cc
+++ b/gcc/lto-cgraph.cc
@@ -1587,3 +1587,5 @@ input_cgraph_1 (struct lto_file_decl_data *file_data,

+#ifndef ACCEL_COMPILER
   lto_input_toplevel_asms (file_data, file_data->order_base);
+#endif

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Not
--- gcc/lto-cgraph.cc   2023-01-09 13:30:47.105343886 +0100
+++ gcc/lto-cgraph.cc   2023-05-11 23:05:54.266991677 +0200
@@ -1023,7 +1023,8 @@ output_symtab (void)
   if (!asm_nodes_output)
     {
       asm_nodes_output = true;
-      lto_output_toplevel_asms ();
+      if (!lto_stream_offload_p)
+       lto_output_toplevel_asms ();
     }

   output_refs (encoder);
instead?
I think that matches conceptually the
+#ifndef ACCEL_COMPILER
   lto_input_toplevel_asms (file_data, file_data->order_base);
+#endif
change.

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #8 from Tobias Burnus <burnus at gcc dot gnu.org> ---
First, I want to remark that inside lto_output_toplevel_asms, 
lto_stream_offload_p is true even with -flto.

However, the following is for output_symtab:

(In reply to Jakub Jelinek from comment #7)
> Not
> --- gcc/lto-cgraph.cc	2023-01-09 13:30:47.105343886 +0100
> +++ gcc/lto-cgraph.cc	2023-05-11 23:05:54.266991677 +0200
> @@ -1023,7 +1023,8 @@ output_symtab (void)
>    if (!asm_nodes_output)
>      {
>        asm_nodes_output = true;
> -      lto_output_toplevel_asms ();
> +      if (!lto_stream_offload_p)
> +	lto_output_toplevel_asms ();

This seems to work, kind of. The only case (I found) where it fails is for:
  -flto -fwpa
Because in that case, lto_stream_offload_p == 1. While there is a diagnostic
(warning), it still duly continues and uses the wrong value. The diagnostic is:

cc1plus: warning: command-line option '-fwpa' is valid for LTO but not for C++

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |13.2
           Priority|P3                          |P2

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Tobias Burnus <burnus@gcc.gnu.org>:

https://gcc.gnu.org/g:a835f046cdf017b9e8ad5576df4f10daaf8420d0

commit r14-793-ga835f046cdf017b9e8ad5576df4f10daaf8420d0
Author: Tobias Burnus <tobias@codesourcery.com>
Date:   Fri May 12 16:27:40 2023 +0200

    LTO: Fix writing of toplevel asm with offloading [PR109816]

    When offloading was enabled, top-level 'asm' were added to the offloading
    section, confusing assemblers which did not support the syntax.
Additionally,
    with offloading and -flto, the top-level assembler code did not end up
    in the host files.

    As r14-321-g9a41d2cdbcd added top-level 'asm' to one libstdc++ header file,
    the issue became more apparent, causing fails with nvptx for some
    C++ testcases.

            PR libstdc++/109816

    gcc/ChangeLog:

            * lto-cgraph.cc (output_symtab): Guard lto_output_toplevel_asms by
            '!lto_stream_offload_p'.

    libgomp/ChangeLog:

            * testsuite/libgomp.c++/target-map-class-1.C: New test.
            * testsuite/libgomp.c++/target-map-class-2.C: New test.

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-13 branch has been updated by Tobias Burnus
<burnus@gcc.gnu.org>:

https://gcc.gnu.org/g:7fb7d49b3c7129fa8a6b500f260abc642929361b

commit r13-7335-g7fb7d49b3c7129fa8a6b500f260abc642929361b
Author: Tobias Burnus <tobias@codesourcery.com>
Date:   Fri May 12 16:27:40 2023 +0200

    LTO: Fix writing of toplevel asm with offloading [PR109816]

    When offloading was enabled, top-level 'asm' were added to the offloading
    section, confusing assemblers which did not support the syntax.
Additionally,
    with offloading and -flto, the top-level assembler code did not end up
    in the host files.

    As r14-321-g9a41d2cdbcd added top-level 'asm' to one libstdc++ header file,
    the issue became more apparent, causing fails with nvptx for some
    C++ testcases.

            PR libstdc++/109816

    gcc/ChangeLog:

            * lto-cgraph.cc (output_symtab): Guard lto_output_toplevel_asms by
            '!lto_stream_offload_p'.

    libgomp/ChangeLog:

            * testsuite/libgomp.c++/target-map-class-1.C: New test.
            * testsuite/libgomp.c++/target-map-class-2.C: New test.

    (cherry picked from commit a835f046cdf017b9e8ad5576df4f10daaf8420d0)

[Bug libstdc++/109816] [13/14 Regression] r14-321 fix for PR108969 breaks NVPTX offloading due to __asm (".globl _ZSt21ios_base_library_initv")

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109816

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #11 from Tobias Burnus <burnus at gcc dot gnu.org> ---
FIXED on mainline (GCC 14) and on the GCC 13 branch (13.1.1).

The issue also exists on previous branches, but only since GCC 13's
'asm' use on 'iostream' [PR108969], it became widely exposed to users.

Hence, the fix was only applied to the such affected branches.