Bug 43622 - Incomplete C++ library support for __float128
Summary: Incomplete C++ library support for __float128
Status: RESOLVED FIXED
Alias: None
Product: gcc
Classification: Unclassified
Component: libstdc++ (show other bugs)
Version: 4.5.0
: P3 enhancement
Target Milestone: ---
Assignee: Benjamin Kosnik
URL:
Keywords:
: 40855 (view as bug list)
Depends on:
Blocks:
 
Reported: 2010-04-01 23:26 UTC by Roman Kononov
Modified: 2023-05-16 19:43 UTC (History)
10 users (show)

See Also:
Host: x86_64-unknown-linux-gnu
Target: x86_64-unknown-linux-gnu
Build: x86_64-unknown-linux-gnu
Known to work:
Known to fail:
Last reconfirmed: 2015-04-04 00:00:00


Attachments
typeinfo exports for float128 (1.79 KB, application/octet-stream)
2011-02-24 18:54 UTC, Benjamin Kosnik
Details
This test-case shows how typeinfo for non-complex 128-bit types DOES NOT work, but typeinfo for complex 128-bit types DOES work. (784 bytes, text/x-c++src)
2011-04-14 15:56 UTC, Steve Ward
Details
potential export fix (1.99 KB, patch)
2014-04-22 17:55 UTC, Marc Glisse
Details | Diff

Note You need to log in before you can comment on or make changes to this bug.
Description Roman Kononov 2010-04-01 23:26:22 UTC
$ cat tf-test.cpp
#include <typeinfo>
typedef float type __attribute__((__mode__(TF)));
int main() { return *(typeid(type).name()); }

$ g++ tf-test.cpp
/tmp/cc8RAh4o.o: In function `main':
tf-test.cpp:(.text+0x5): undefined reference to `typeinfo for __float128'
collect2: ld returned 1 exit status

$ cat ti-test.cpp
#include <typeinfo>
typedef int type __attribute__((__mode__(TI)));
int main() { return *(typeid(type).name()); }

$ g++ ti-test.cpp
/tmp/ccI1nTTK.o: In function `main':
ti-test.cpp:(.text+0x5): undefined reference to `typeinfo for __int128'
collect2: ld returned 1 exit status

$ g++ --version | head -1
g++ (GCC) 4.5.0 20100401 (experimental)
Comment 1 Benjamin Kosnik 2011-02-24 07:33:56 UTC
Mine.
Comment 2 jsm-csl@polyomino.org.uk 2011-02-24 15:23:33 UTC
This seems related to bug 40855.  See also 
<http://gcc.gnu.org/ml/gcc-patches/2009-11/msg00652.html> and the rest of 
that thread.  libstdc++ support for extended integer and floating-point 
types covers more than typeinfo; see what I said at 
<http://gcc.gnu.org/ml/gcc-patches/2010-05/msg01912.html> and the 
followups thereto.  There's numeric_limits (see bug 40856), and I/O 
support (with the same issues as for libgfortran about avoiding 
libquadmath in static links if the functionality isn't used in a 
particular program), and mathematical functions, and maybe more - 
certainly most of that would not be suitable for 4.6, however.
Comment 3 Benjamin Kosnik 2011-02-24 18:53:08 UTC
Expecting this as exported as fundamental_type_info, see in emit_support_tinfos via rtti.c:1461:

  static tree *const fundamentals[] =
  {
    &void_type_node,
    &boolean_type_node,
    &wchar_type_node, &char16_type_node, &char32_type_node,
    &char_type_node, &signed_char_type_node, &unsigned_char_type_node,
    &short_integer_type_node, &short_unsigned_type_node,
    &integer_type_node, &unsigned_type_node,
    &long_integer_type_node, &long_unsigned_type_node,
    &long_long_integer_type_node, &long_long_unsigned_type_node,
    &int128_integer_type_node, &int128_unsigned_type_node,
    &float_type_node, &double_type_node, &long_double_type_node,
    &dfloat32_type_node, &dfloat64_type_node, &dfloat128_type_node,
    &nullptr_type_node,
    0
  };
 
Which should take care of this, given the libstdc++ ver patch to export the symbols. However, none is emitted when building libsupc++/fundamental_type_info.o.

Ouch.
Comment 4 Benjamin Kosnik 2011-02-24 18:54:06 UTC
Created attachment 23457 [details]
typeinfo exports for float128
Comment 5 Benjamin Kosnik 2011-02-24 19:09:33 UTC
RE comment 2. 

Yes, agreed full support is a ways off. However, typeinfo support is ostensibly already there, just not emitted. This is a bug, and not something users can work around. IMHO it should be fixed, and since there are already 4.6.0 typeinfo exports this seems like an opportune time to do it.
Comment 6 Jakub Jelinek 2011-02-24 19:13:16 UTC
I guess #c4 patch will break ppc/s390 etc.
config/os/gnu-linux/ldbl-extra.ver already has:
CXXABI_LDBL_1.3 {
  _ZT[IS]g;
  _ZT[IS]Pg;
  _ZT[IS]PKg;
};

So, IMNSHO you don't want to add something like that to gnu.ver, but have
a special *.ver file for those 3 targets (i?86/x86_64/ia64) that have __float128.
Comment 7 Benjamin Kosnik 2011-02-24 19:28:02 UTC
ack, I mis-read rtti.c, these are the decimal exports, ie decimal128  not float128.

Jakub you are correct these are exports as LD symbols via ld versioning via inclusion. I missed that!

%find . -type f | xargs grep _ZTIg
./powerpc-linux-gnu/baseline_symbols.txt:OBJECT:8:_ZTIg@@CXXABI_LDBL_1.3
./sparc-linux-gnu/baseline_symbols.txt:OBJECT:8:_ZTIg@@CXXABI_LDBL_1.3
./powerpc64-linux-gnu/32/baseline_symbols.txt:OBJECT:8:_ZTIg@@CXXABI_LDBL_1.3
./powerpc64-linux-gnu/baseline_symbols.txt:OBJECT:16:_ZTIg@@CXXABI_LDBL_1.3
./s390-linux-gnu/baseline_symbols.txt:OBJECT:8:_ZTIg@@CXXABI_LDBL_1.3
./s390x-linux-gnu/baseline_symbols.txt:OBJECT:16:_ZTIg@@CXXABI_LDBL_1.3
./alpha-linux-gnu/baseline_symbols.txt:OBJECT:16:_ZTIg@@CXXABI_LDBL_1.3

sparc may be questionable in reality as it is inferred on alpha.
Comment 8 Jakub Jelinek 2011-03-25 19:52:22 UTC
GCC 4.6.0 is being released, adjusting target milestone.
Comment 9 Steve Ward 2011-04-14 15:56:12 UTC
Created attachment 23982 [details]
This test-case shows how typeinfo for non-complex 128-bit types DOES NOT work, but typeinfo for complex 128-bit types DOES work.

$ g++-4.5 --version
g++-4.5 (Ubuntu/Linaro 4.5.1-7ubuntu2) 4.5.1
Copyright (C) 2010 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ uname --all
Linux dev8 2.6.35-28-generic #50-Ubuntu SMP Fri Mar 18 18:42:20 UTC 2011 x86_64 GNU/Linux

$ g++-4.5 gcc_bug_43622.cpp && ./a.out
complex_int128 = Cn
complex_uint128 = Co
complex_binary128 = Cg
Comment 10 Paolo Carlini 2011-09-25 15:04:17 UTC
*** Bug 40855 has been marked as a duplicate of this bug. ***
Comment 11 Steve Ward 2012-10-21 22:05:35 UTC
This problem still exists in g++ 4.7.2.
Comment 12 Paul A. Bristow 2014-03-02 10:41:15 UTC
This still exists at 4.8.2 and is a *showstopper* for running the Boost.Math library at all the available precisions, up to 128-bit precision where available.

typeid(type).name() fails with:

undefined reference to `typeinfo for __float128'

We can check that the library seems to work OK by ugly hacking of error handling and a few examples of test code (out of the hundreds of tests), but we absolutely need this before it can be fully tested at 128-bit precision and released.

Getting this library to pass is part of a demonstration of the proposed C++ and C library additions by Christopher Kormanyos and John Maddock

Floating-Point Typedefs Having Specified Widths - N3626

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/n3626.pdf

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2004/n1703.pdf

Everything is working to provide full C++ 128-bit floating-point - apart from this :-( 

So we are very keen to have a fix.
Comment 13 Marc Glisse 2014-03-02 12:15:32 UTC
It looks like emit_support_tinfos (rtti.c) should go through registered_builtin_types (hidden in c-common.c) in addition to the hardcoded fundamentals list.
Comment 14 Marc Glisse 2014-03-02 16:11:26 UTC
Updating the title: it seems to me that typeinfo for __int128 has been available for a while, it is only __float128 that is missing.
Comment 15 Marc Glisse 2014-04-22 16:45:17 UTC
Author: glisse
Date: Tue Apr 22 16:44:46 2014
New Revision: 209652

URL: http://gcc.gnu.org/viewcvs?rev=209652&root=gcc&view=rev
Log:
2014-04-22  Marc Glisse  <marc.glisse@inria.fr>

	PR libstdc++/43622
gcc/c-family/
	* c-common.c (registered_builtin_types): Make non-static.
	* c-common.h (registered_builtin_types): Declare.
gcc/cp/
	* rtti.c (emit_support_tinfo_1): New function, extracted from
	emit_support_tinfos.
	(emit_support_tinfos): Call it and iterate on registered_builtin_types.
libstdc++-v3/
	* config/abi/pre/gnu.ver (CXXABI_1.3.9): New version, new symbols.
	* config/abi/pre/gnu-versioned-namespace.ver: New symbols.
	* config/abi/post/x86_64-linux-gnu/baseline_symbols.txt: Likewise.



Modified:
    trunk/gcc/c-family/ChangeLog
    trunk/gcc/c-family/c-common.c
    trunk/gcc/c-family/c-common.h
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/rtti.c
    trunk/libstdc++-v3/ChangeLog
    trunk/libstdc++-v3/config/abi/post/x86_64-linux-gnu/baseline_symbols.txt
    trunk/libstdc++-v3/config/abi/pre/gnu-versioned-namespace.ver
    trunk/libstdc++-v3/config/abi/pre/gnu.ver
Comment 16 Marc Glisse 2014-04-22 17:06:05 UTC
(In reply to Jakub Jelinek from comment #6)
> I guess #c4 patch will break ppc/s390 etc.
> config/os/gnu-linux/ldbl-extra.ver already has:
> CXXABI_LDBL_1.3 {
>   _ZT[IS]g;
>   _ZT[IS]Pg;
>   _ZT[IS]PKg;
> };
> 
> So, IMNSHO you don't want to add something like that to gnu.ver, but have
> a special *.ver file for those 3 targets (i?86/x86_64/ia64) that have
> __float128.

Er, oups, I was going to close the bug, but apparently I overlooked this comment when writing the patch, sorry :-(

Should I revert right away?

It looks like I need to create a file config/?/?/float128.ver and find the right incantation in configure to add it to port_specific_symbol_files for those 3 targets?
Comment 17 Jonathan Wakely 2014-04-22 17:17:36 UTC
(In reply to Marc Glisse from comment #16)
> Should I revert right away?

If it doesn't break bootstrap for the ldbl targets (only causes test failures) then I would say reverting it is not necessary if you will be able to work on a fix fairly quickly.

> It looks like I need to create a file config/?/?/float128.ver and find the
> right incantation in configure to add it to port_specific_symbol_files for
> those 3 targets?

I honestly don't know, but that sounds reasonable.
Comment 18 Marc Glisse 2014-04-22 17:25:19 UTC
The easiest would be to move CXXABI_LDBL_1.3 to gnu.ver and use that for the new __float128 symbols on x86, but I'm sure that would be considered too hacky.
Comment 19 Marc Glisse 2014-04-22 17:55:18 UTC
Created attachment 32654 [details]
potential export fix

I am currently testing the attached.
Comment 20 Marc Glisse 2014-04-24 13:59:08 UTC
Author: glisse
Date: Thu Apr 24 13:58:36 2014
New Revision: 209748

URL: http://gcc.gnu.org/viewcvs?rev=209748&root=gcc&view=rev
Log:
2014-04-24  Marc Glisse  <marc.glisse@inria.fr>

	PR libstdc++/43622
gcc/cp/
	* rtti.c (emit_support_tinfos): Do not iterate on
	registered_builtin_types (partial revert).
libstdc++/
	* config/abi/pre/gnu.ver (CXXABI_1.3.9): Remove __float128 symbols.
	* config/abi/pre/gnu-versioned-namespace.ver: Likewise.
	* config/abi/post/x86_64-linux-gnu/baseline_symbols.txt: Update.


Modified:
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/rtti.c
    trunk/libstdc++-v3/ChangeLog
    trunk/libstdc++-v3/config/abi/post/x86_64-linux-gnu/baseline_symbols.txt
    trunk/libstdc++-v3/config/abi/pre/gnu-versioned-namespace.ver
    trunk/libstdc++-v3/config/abi/pre/gnu.ver
Comment 21 Marc Glisse 2014-11-18 20:21:25 UTC
Author: glisse
Date: Tue Nov 18 20:20:53 2014
New Revision: 217735

URL: https://gcc.gnu.org/viewcvs?rev=217735&root=gcc&view=rev
Log:
2014-11-18  Marc Glisse  <marc.glisse@inria.fr>

	PR libstdc++/43622
gcc/cp/
	* rtti.c (emit_support_tinfos): Handle __float128.
libstdc++-v3/
	* config/abi/pre/float128.ver: New file.
	* configure.ac: Use float128.ver when relevant.
	* configure: Regenerate.
	* testsuite/util/testsuite_abi.cc (check_version): Accept new
	CXXABI_FLOAT128 version.


Added:
    trunk/libstdc++-v3/config/abi/pre/float128.ver
Modified:
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/rtti.c
    trunk/libstdc++-v3/ChangeLog
    trunk/libstdc++-v3/configure
    trunk/libstdc++-v3/configure.ac
    trunk/libstdc++-v3/testsuite/util/testsuite_abi.cc
Comment 22 Marc Glisse 2014-11-18 20:42:04 UTC
__float128 is still missing a specialization of numeric_limits.
Comment 23 jsm-csl@polyomino.org.uk 2014-11-18 23:50:50 UTC
On Tue, 18 Nov 2014, glisse at gcc dot gnu.org wrote:

> __float128 is still missing a specialization of numeric_limits.

Fully supporting an extended type (whether floating-point, or one like 
__int128 that mostly acts like an integer type) includes (as I noted in 
one of the other past bug reports on such issues, bug 50441 (fixed)) 
typeinfo, numeric_limits (maybe with associated built-in macros), I/O and 
mathematical functions in general.

But you may not want libstdc++ to depend on libquadmath, which may limit 
what you can support for __float128.  (The ideal would be full C support 
for the relevant parts of TS 18661 in GCC and glibc, which would give you 
library support in libc and libm that could then be used from libstdc++ 
when used with glibc.)

I think it would be reasonable to define built-in __FLT128_*__ macros for 
__float128 similar to those defined for other floating-point types (you'd 
then make quadmath.h use those instead of hardcoding the values directly) 
- and you could then use those macros in numeric_limits.  That naming is 
in line with DTS 18661-3 defining e.g. FLT128_MANT_DIG as public names for 
macros relating to _Float128.
Comment 24 John Maddock 2014-11-20 17:11:56 UTC
(In reply to joseph@codesourcery.com from comment #23)
> On Tue, 18 Nov 2014, glisse at gcc dot gnu.org wrote:
> 
> > __float128 is still missing a specialization of numeric_limits.
> 
> Fully supporting an extended type (whether floating-point, or one like 
> __int128 that mostly acts like an integer type) includes (as I noted in 
> one of the other past bug reports on such issues, bug 50441 (fixed)) 
> typeinfo, numeric_limits (maybe with associated built-in macros), I/O and 
> mathematical functions in general.
> 
> But you may not want libstdc++ to depend on libquadmath, which may limit 
> what you can support for __float128.  (The ideal would be full C support 
> for the relevant parts of TS 18661 in GCC and glibc, which would give you 
> library support in libc and libm that could then be used from libstdc++ 
> when used with glibc.)
> 
> I think it would be reasonable to define built-in __FLT128_*__ macros for 
> __float128 similar to those defined for other floating-point types (you'd 
> then make quadmath.h use those instead of hardcoding the values directly) 
> - and you could then use those macros in numeric_limits.  That naming is 
> in line with DTS 18661-3 defining e.g. FLT128_MANT_DIG as public names for 
> macros relating to _Float128.

+1 on supporting numeric_limits.

IMO while it would be nice to provide full standard lib support, there are levels of importance here.  typeinfo was essential because the user cannot add that support themselves.  numeric_limits and iostream support come next in importance, and of course std::whatever overloads for the <cmath> functions would be nice but come lower down I guess.

I understand the desire to limit the dependency on libquadmath - perhaps first class std lib support should be added to <quadmath.h> rather than <limits>, <cmath> etc if this is an overarching concern?  That said, implementating numeric_limits for that type is trivial, and this is obviously a built in type, not a pure library extension so there is a lot to be said for providing this support without the need to include additional headers.

While we're opening cans of worms.... intmax_t should clearly be __int128... just saying!
Comment 25 jsm-csl@polyomino.org.uk 2014-11-20 22:26:26 UTC
On Thu, 20 Nov 2014, john at johnmaddock dot co.uk wrote:

> While we're opening cans of worms.... intmax_t should clearly be __int128...
> just saying!

Existing ABIs where intmax_t in libc is 64-bit are why __int128 is what I 
call a sui generis extended type, not an integer type.
Comment 26 John Maddock 2014-11-21 12:18:41 UTC
(In reply to joseph@codesourcery.com from comment #25)
> On Thu, 20 Nov 2014, john at johnmaddock dot co.uk wrote:
> 
> > While we're opening cans of worms.... intmax_t should clearly be __int128...
> > just saying!
> 
> Existing ABIs where intmax_t in libc is 64-bit are why __int128 is what I 
> call a sui generis extended type, not an integer type.

So it's an integer that's not an integer?  I'm sorry but that's just nonesense.  Of course I realise that ABI issues may trump other concerns, but please call a spade a spade!  In any case this is a glibc issue and we're off topic here...
Comment 27 Marc Glisse 2015-04-04 13:36:32 UTC
Renaming since typeinfo is in gcc-5.
Comment 28 Andrew Pinski 2023-05-16 18:19:29 UTC
I suspect this has now been fixed with the additional of full _Float128 support in the C++ front-end and the library work.
Comment 29 Jonathan Wakely 2023-05-16 18:47:45 UTC
(In reply to Marc Glisse from comment #22)
> __float128 is still missing a specialization of numeric_limits.

We're still missing that, but I created PR libstdc++/104772 for it. I don't think we need full <cmath> and I/O support for __float128, so let's close this one.
Comment 30 Jonathan Wakely 2023-05-16 18:49:09 UTC
(In reply to John Maddock from comment #26)
> (In reply to joseph@codesourcery.com from comment #25)
> > On Thu, 20 Nov 2014, john at johnmaddock dot co.uk wrote:
> > 
> > > While we're opening cans of worms.... intmax_t should clearly be __int128...
> > > just saying!
> > 
> > Existing ABIs where intmax_t in libc is 64-bit are why __int128 is what I 
> > call a sui generis extended type, not an integer type.
> 
> So it's an integer that's not an integer?  I'm sorry but that's just
> nonesense.  Of course I realise that ABI issues may trump other concerns,
> but please call a spade a spade!  In any case this is a glibc issue and
> we're off topic here...

With changes to the definition of intmax_t in C2x (and C++23) that problem is gone. __int128 can be an extended integer type, and intmax_t doesn't need to change, so there's no ABI problem. Order is restored to the galaxy.
Comment 31 jsm-csl@polyomino.org.uk 2023-05-16 19:33:31 UTC
It can be an extended integer type in C2x, but then stdint.h would be 
required to have int128_t / uint128_t / int_least128_t / uint_least128_t 
typedefs, and integer constant suffixes would be needed for the 
corresponding macros INT128_C / UINT128_C (and the other stdint.h macros 
for the types would need to be defined as well), and printf/scanf support 
would be required as well.
Comment 32 Jonathan Wakely 2023-05-16 19:43:47 UTC
Hmm, yes. Well I think we can at least make std::is_integral<__int128> true, which will remove once source of surprises for users.