libstdc++ "freestanding" ('--disable-hosted-libstdcxx') with '-fno-rtti', '-fno-exceptions': 'libstdc++-v3/libsupc++/tinfo.cc'

libstdc++ "freestanding" ('--disable-hosted-libstdcxx') with '-fno-rtti', '-fno-exceptions': 'libstdc++-v3/libsupc++/tinfo.cc'

[Thomas Schwinge]

Hi!

In context of <https://gcc.gnu.org/PR101544> '[OpenMP][AMDGCN][nvptx]
C++ offloading: unresolved _Znwm = "operator new(unsigned long)"'
I'm looking into building GCN, nvptx offloading libstdc++ "freestanding"
('--disable-hosted-libstdcxx') with '-fno-rtti', '-fno-exceptions'.
(I've basically got these things wired up; details to be shared later.)

While there is some experimental/incomplete/not-to-be-relied-on support
for PTX symbol aliases, we're currently generally running into
"error: alias definitions not supported in this configuration"
for certain GCC/C++-front-end-generated code, for example:
"complete object destructor" aliasing "base object destructor"
(<https://itanium-cxx-abi.github.io/cxx-abi/abi.html#mangling-special-ctor-dtor>
-- per my understanding, at least).  ;-)

It'll ultimately be possible to implement more complete symbol aliasing
support for nvptx, but I found that this specific issue actually is due
to a nvptx back end mis-configuration; fixed with adequate
'TARGET_USE_LOCAL_THUNK_ALIAS_P' and/or 'TARGET_SUPPORTS_ALIASES'
definitions (details to be shared later).

However, we then still run into:

    [...]/libstdc++-v3/libsupc++/tinfo.cc:55:1: error: alias definitions not supported in this configuration
       55 | std::type_info::__equal (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
          | ^~~
    make[4]: *** [Makefile:777: tinfo.lo] Error 1

That's 'libstdc++-v3/libsupc++/tinfo.cc':

     1 // Methods for type_info for -*- C++ -*- Run Time Type Identification.
    [...]
    39 // We can't rely on common symbols being shared between shared objects.
    40 bool std::type_info::
    41 operator== (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
    42 {
    43 #if __GXX_MERGED_TYPEINFO_NAMES
    44   return name () == arg.name ();
    45 #else
    46   /* The name() method will strip any leading '*' prefix. Therefore
    47      take care to look at __name rather than name() when looking for
    48      the "pointer" prefix.  */
    49   return (&arg == this)
    50     || (__name[0] != '*' && (__builtin_strcmp (name (), arg.name ()) == 0));
    51 #endif
    52 }
    53
    54 bool
    55 std::type_info::__equal (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
    56 __attribute__((alias("_ZNKSt9type_infoeqERKS_")));
    57 #endif
    [...]

..., so there's a manual alias from the line 55 function to the line 41
function (if I got that right).

Now we may surely address that in some different way, but I do wonder if
we need this whole file at all, given this is
"Methods for type_info for -*- C++ -*- Run Time Type Identification",
RTTI, and we're generally assuming '-fno-rtti'?  Do certain things simply
need to be '#if'-conditionalized, or similar, for example?  I've not yet
looked into the details; hoping that's maybe easy for you to answer?


Grüße
 Thomas
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955

Re: libstdc++ "freestanding" ('--disable-hosted-libstdcxx') with '-fno-rtti', '-fno-exceptions': 'libstdc++-v3/libsupc++/tinfo.cc'

[Jonathan Wakely]

On Thu, 14 Jul 2022 at 19:14, Thomas Schwinge <thomas@codesourcery.com> wrote:
>
> Hi!
>
> In context of <https://gcc.gnu.org/PR101544> '[OpenMP][AMDGCN][nvptx]
> C++ offloading: unresolved _Znwm = "operator new(unsigned long)"'
> I'm looking into building GCN, nvptx offloading libstdc++ "freestanding"
> ('--disable-hosted-libstdcxx') with '-fno-rtti', '-fno-exceptions'.
> (I've basically got these things wired up; details to be shared later.)
>
> While there is some experimental/incomplete/not-to-be-relied-on support
> for PTX symbol aliases, we're currently generally running into
> "error: alias definitions not supported in this configuration"
> for certain GCC/C++-front-end-generated code, for example:
> "complete object destructor" aliasing "base object destructor"
> (<https://itanium-cxx-abi.github.io/cxx-abi/abi.html#mangling-special-ctor-dtor>
> -- per my understanding, at least).  ;-)
>
> It'll ultimately be possible to implement more complete symbol aliasing
> support for nvptx, but I found that this specific issue actually is due
> to a nvptx back end mis-configuration; fixed with adequate
> 'TARGET_USE_LOCAL_THUNK_ALIAS_P' and/or 'TARGET_SUPPORTS_ALIASES'
> definitions (details to be shared later).
>
> However, we then still run into:
>
>     [...]/libstdc++-v3/libsupc++/tinfo.cc:55:1: error: alias definitions not supported in this configuration
>        55 | std::type_info::__equal (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
>           | ^~~
>     make[4]: *** [Makefile:777: tinfo.lo] Error 1
>
> That's 'libstdc++-v3/libsupc++/tinfo.cc':
>
>      1 // Methods for type_info for -*- C++ -*- Run Time Type Identification.
>     [...]
>     39 // We can't rely on common symbols being shared between shared objects.
>     40 bool std::type_info::
>     41 operator== (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
>     42 {
>     43 #if __GXX_MERGED_TYPEINFO_NAMES
>     44   return name () == arg.name ();
>     45 #else
>     46   /* The name() method will strip any leading '*' prefix. Therefore
>     47      take care to look at __name rather than name() when looking for
>     48      the "pointer" prefix.  */
>     49   return (&arg == this)
>     50     || (__name[0] != '*' && (__builtin_strcmp (name (), arg.name ()) == 0));
>     51 #endif
>     52 }
>     53
>     54 bool
>     55 std::type_info::__equal (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
>     56 __attribute__((alias("_ZNKSt9type_infoeqERKS_")));
>     57 #endif
>     [...]
>
> ..., so there's a manual alias from the line 55 function to the line 41
> function (if I got that right).

That's only there for backwards compatibility on ARM EABI and other
targets that don't define __GXX_TYPEINFO_EQUALITY_INLINE==1.

My suggestion would be to define that macro for the target.

>
> Now we may surely address that in some different way, but I do wonder if
> we need this whole file at all, given this is
> "Methods for type_info for -*- C++ -*- Run Time Type Identification",
> RTTI, and we're generally assuming '-fno-rtti'?  Do certain things simply
> need to be '#if'-conditionalized, or similar, for example?  I've not yet
> looked into the details; hoping that's maybe easy for you to answer?

Again, is -fno-rtti globally true for all code for this target? Not
just while building libstdc++?

If the answer is yes, then you don't need the symbols in this file.

nvptx: Use the usual '#define MAKE_DECL_ONE_ONLY(DECL) (DECL_WEAK (DECL) = 1)' (was: libstdc++ "freestanding" ('--disable-hosted-libstdcxx') with '-fno-rtti', '-fno-exceptions': 'libstdc++-v3/libsupc++/tinfo.cc')

[Thomas Schwinge]

[-- Attachment #1: Type: text/plain, Size: 2624 bytes --]

Hi!

On 2022-07-14T19:39:40+0100, Jonathan Wakely <jwakely@redhat.com> wrote:
> On Thu, 14 Jul 2022 at 19:14, Thomas Schwinge <thomas@codesourcery.com> wrote:
>> I'm looking into building GCN, nvptx offloading libstdc++ [...]

>> While there is some experimental/incomplete/not-to-be-relied-on support
>> for PTX symbol aliases, we're currently generally running into
>> "error: alias definitions not supported in this configuration"

>>     [...]/libstdc++-v3/libsupc++/tinfo.cc:55:1: error: alias definitions not supported in this configuration
>>        55 | std::type_info::__equal (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
>>           | ^~~
>>     make[4]: *** [Makefile:777: tinfo.lo] Error 1
>>
>> That's 'libstdc++-v3/libsupc++/tinfo.cc':
>>
>>      1 // Methods for type_info for -*- C++ -*- Run Time Type Identification.
>>     [...]
>>     39 // We can't rely on common symbols being shared between shared objects.
>>     40 bool std::type_info::
>>     41 operator== (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
>>     42 {
>>     43 #if __GXX_MERGED_TYPEINFO_NAMES
>>     44   return name () == arg.name ();
>>     45 #else
>>     46   /* The name() method will strip any leading '*' prefix. Therefore
>>     47      take care to look at __name rather than name() when looking for
>>     48      the "pointer" prefix.  */
>>     49   return (&arg == this)
>>     50     || (__name[0] != '*' && (__builtin_strcmp (name (), arg.name ()) == 0));
>>     51 #endif
>>     52 }
>>     53
>>     54 bool
>>     55 std::type_info::__equal (const std::type_info& arg) const _GLIBCXX_NOEXCEPT
>>     56 __attribute__((alias("_ZNKSt9type_infoeqERKS_")));
>>     57 #endif
>>     [...]
>>
>> ..., so there's a manual alias from the line 55 function to the line 41
>> function (if I got that right).
>
> That's only there for backwards compatibility on ARM EABI and other
> targets that don't define __GXX_TYPEINFO_EQUALITY_INLINE==1.
>
> My suggestion would be to define that macro for the target.

Thanks, that did put me on the right track -- we shall define it this
way; indirectly.

Pushed to master branch commit 9837f62f066db532c9db6df38ccf2653d0c3a960
"nvptx: Use the usual '#define MAKE_DECL_ONE_ONLY(DECL) (DECL_WEAK (DECL) = 1)'",
see attached.


Grüße
 Thomas


-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: 0001-nvptx-Use-the-usual-define-MAKE_DECL_ONE_ONLY-DECL-D.patch --]
[-- Type: text/x-diff, Size: 2765 bytes --]

From 9837f62f066db532c9db6df38ccf2653d0c3a960 Mon Sep 17 00:00:00 2001
From: Thomas Schwinge <thomas@codesourcery.com>
Date: Thu, 14 Jul 2022 23:22:35 +0200
Subject: [PATCH] nvptx: Use the usual '#define MAKE_DECL_ONE_ONLY(DECL)
 (DECL_WEAK (DECL) = 1)'

With this 'MAKE_DECL_ONE_ONLY' definition, we get 'SUPPORTS_ONE_ONLY', and thus
'__GXX_WEAK__', and thus '__GXX_TYPEINFO_EQUALITY_INLINE'.  This unblocks build
of 'libstdc++-v3/libsupc++/tinfo.cc', which otherwise depends on symbol alias
support, which GCC/nvptx doesn't generally provide.  Also, this gets us a
number of FAIL -> PASS progressions in the test suite.

Given that GCC/nvptx support for weak symbols isn't complete, we also get a few
more of the already-known
'error: PTX does not support weak declarations (only weak definitions)':

    [-PASS:-]{+FAIL:+} g++.old-deja/g++.other/crash11.C  -std=c++14 (test for excess errors)
    [-PASS:-]{+FAIL:+} g++.old-deja/g++.other/crash11.C  -std=c++17 (test for excess errors)
    [-PASS:-]{+FAIL:+} g++.old-deja/g++.other/crash11.C  -std=c++20 (test for excess errors)
    [-PASS:-]{+FAIL:+} g++.old-deja/g++.other/crash11.C  -std=c++98 (test for excess errors)

    [-PASS:-]{+FAIL:+} g++.old-deja/g++.pt/crash29.C  -std=c++14 (test for excess errors)
    [-PASS:-]{+FAIL:+} g++.old-deja/g++.pt/crash29.C  -std=c++17 (test for excess errors)
    [-PASS:-]{+FAIL:+} g++.old-deja/g++.pt/crash29.C  -std=c++20 (test for excess errors)
    [-PASS:-]{+FAIL:+} g++.old-deja/g++.pt/crash29.C  -std=c++98 (test for excess errors)

    [-PASS:-]{+FAIL:+} 23_containers/map/56613.cc  -std=gnu++17 (test for excess errors)

... as well as one more of the already-known
'sorry, unimplemented: target cannot support nonlocal goto':

    PASS: g++.dg/tree-ssa/pr22488.C  -std=gnu++14 (test for excess errors)
    PASS: g++.dg/tree-ssa/pr22488.C  -std=gnu++17 (test for excess errors)
    PASS: g++.dg/tree-ssa/pr22488.C  -std=gnu++20 (test for excess errors)
    [-PASS:-]{+FAIL:+} g++.dg/tree-ssa/pr22488.C  -std=gnu++98 (test for excess errors)

We shall look into these, later.

	gcc/
	* config/nvptx/nvptx.h (MAKE_DECL_ONE_ONLY): Define.
---
 gcc/config/nvptx/nvptx.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/gcc/config/nvptx/nvptx.h b/gcc/config/nvptx/nvptx.h
index 129427e5654..407fd70f06a 100644
--- a/gcc/config/nvptx/nvptx.h
+++ b/gcc/config/nvptx/nvptx.h
@@ -319,6 +319,9 @@ struct GTY(()) machine_function
 
 #define SUPPORTS_WEAK 1
 
+#define MAKE_DECL_ONE_ONLY(DECL) \
+  (DECL_WEAK (DECL) = 1)
+
 /* The documentation states that ASM_OUTPUT_DEF_FROM_DECLS is used in
    preference to ASM_OUTPUT_DEF if the tree nodes are available.  However, we
    need the tree nodes to emit the prototype, so at this point it's not clear
-- 
2.34.1