https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119806
Bug ID: 119806
Summary: OpenACC, OpenMP 'target' offloading vs. C++ 'typeinfo'
parts
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: openacc, openmp, wrong-code
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: tschwinge at gcc dot gnu.org
CC: burnus at gcc dot gnu.org, jakub at gcc dot gnu.org
Target Milestone: ---
Target: GCN, nvptx
We've got cases of C++ OpenACC, OpenMP 'target' offloading code (XFAILed test
case soon to appear), where nvptx offloading linking fails due to:
ptxas ./exceptions-throw-2.xnvptx-none.mkoffload.o, line 23; fatal :
Initial value type mismatch
ptxas fatal : Ptx assembly aborted due to errors
nvptx-as: ptxas returned 255 exit status
nvptx mkoffload: fatal error:
[...]/powerpc64le-unknown-linux-gnu-accel-nvptx-none-gcc returned 1 exit status
It's complaining about:
// BEGIN GLOBAL VAR DEF: _ZTI11MyException
.weak .global .align 8 .u64 _ZTI11MyException[2] = {
generic(_ZTVN10__cxxabiv117__class_type_infoE) + 16, generic(_ZTS11MyException)
};
The players here are:
$ c++filt _ZTI11MyException _ZTVN10__cxxabiv117__class_type_infoE
_ZTS11MyException
typeinfo for MyException
vtable for __cxxabiv1::__class_type_info
typeinfo name for MyException
..., but '_ZTVN10__cxxabiv117__class_type_infoE' and '_ZTS11MyException' aren't
getting declared in this file.
This issue bears some relation to PR119734 "nvptx/C++ vs. '_ZTISt8bad_cast'
('typeinfo for std::bad_cast')", as again, '_ZTI11MyException' is only being
"referenced" within:
/* BEGIN '.gcc_except_table'
[...]
.symbol_ref _ZTI11MyException
END '.gcc_except_table' */
It would be good if there was a way to not emit that '.extern'
'_ZTI11MyException' definition if only (fake-)referenced like this.
Here, '_ZTVN10__cxxabiv117__class_type_infoE' comes from libstdc++ (via
definition in libsupc++'s 'class_type_info.o'), and '_ZTS11MyException' would
come from the user source code -- but isn't getting defined in this file.
In contrast, the respective non-offloading, nvptx target test case (to appear
soon), gets this right, includes the 'vtable for __cxxabiv1::__class_type_info'
declaration:
// BEGIN GLOBAL VAR DECL: _ZTVN10__cxxabiv117__class_type_infoE
.extern .global .align 8 .u64
_ZTVN10__cxxabiv117__class_type_infoE[];
..., and includes the 'typeinfo name for MyException' definition:
// BEGIN GLOBAL VAR DEF: _ZTS11MyException
.weak .global .align 1 .u8 _ZTS11MyException[14] = { 49, 49, 77,
121, 69, 120, 99, 101, 112, 116, 105, 111, 110, 0 };
..., and assembles, links, executes as expected.
I suspect that we simply fail to mark up these C++ 'typeinfo' parts to be
emitted for offloading compilation? (Via 'omp declare target' attribute, set
'offloadable' on 'symtab_node', push into 'offload_vars'?)
Quite similarly, GCN offloading linking fails due to:
ld: error: undefined symbol: typeinfo name for MyException
>>> referenced by
./exceptions-throw-2.xamdgcn-amdhsa.mkoffload.hsaco-exceptions-throw-2.xamdgcn-amdhsa.mkoffload.2.o:(typeinfo
for MyException)
>>> did you mean: typeinfo for MyException
>>> defined in:
./exceptions-throw-2.xamdgcn-amdhsa.mkoffload.hsaco-exceptions-throw-2.xamdgcn-amdhsa.mkoffload.2.o
collect2: error: ld returned 1 exit status
gcn mkoffload: fatal error:
[...]/x86_64-pc-linux-gnu-accel-amdgcn-amdhsa-gcc returned 1 exit status
As above, that's:
.weak _ZTI11MyException
.section
.data.rel.ro._ZTI11MyException,"awG",@progbits,_ZTI11MyException,comdat
.align 8
.type _ZTI11MyException, @object
.size _ZTI11MyException, 16
_ZTI11MyException:
.8byte _ZTVN10__cxxabiv117__class_type_infoE+16
.8byte _ZTS11MyException
..., which also here is only referenced from '.fake_gcc_except_table':
.section .fake_gcc_except_table,"e",@progbits
[...]
.8byte _ZTI11MyException
As above, the players here are:
$ c++filt _ZTI11MyException _ZTVN10__cxxabiv117__class_type_infoE
_ZTS11MyException
typeinfo for MyException
vtable for __cxxabiv1::__class_type_info
typeinfo name for MyException
As above, '_ZTVN10__cxxabiv117__class_type_infoE' comes from libstdc++ (via
definition in libsupc++'s 'class_type_info.o'), and '_ZTS11MyException' would
come from the user source code, but isn't getting defined in this file.
As above, in contrast, the respective non-offloading, GCN target test case (to
appear soon), gets this right, (doesn't need a 'vtable for
__cxxabiv1::__class_type_info' declaration, but) includes the 'typeinfo name
for MyException' definition:
.weak _ZTS11MyException
.section
.rodata._ZTS11MyException,"aG",@progbits,_ZTS11MyException,comdat
.align 16
.type _ZTS11MyException, @object
.size _ZTS11MyException, 14
_ZTS11MyException:
.string "11MyException"
..., and assembles, links, executes as expected.
So I suspect it's the very same underlying issue.
