It works if you remove the parentheses, which also makes the code valid C++ (currently it relies on a GNU extension that allows a variable array size for a parenthesised new expression, you get a warning with -pedantic or -Wvla)

It seems that there's some interaction between the C++11 lambda parsing and the GNU extension

Hello, I have something similar in the OpenMM code (https://simtk.org/home/openmm), with no ()s following the "new" as in


template <typename T>
struct CUDAStream : public SoADeviceObject
{
    unsigned int    _length;
    unsigned int    _subStreams;      # line 68
    unsigned int    _stride;
    T**             _pSysStream;
    T**             _pDevStream;
    T*              _pSysData;
    T*              _pDevData;
    std::string     _name;
    CUDAStream(int length, int subStreams = 1, std::string name="");
    CUDAStream(unsigned int length, unsigned int subStreams = 1, std::string name="");
    CUDAStream(unsigned int length, int subStreams = 1, std::string name="");
    CUDAStream(int length, unsigned int subStreams = 1, std::string name="");
    virtual ~CUDAStream();
    void Allocate();
    void Deallocate();
    void Upload();
    void Download();
    void CopyFrom(const CUDAStream<T>& src);
    void Collapse(unsigned int newstreams = 1, unsigned int interleave = 1);
    T& operator[](int index);
};

[...]

template <typename T>
void CUDAStream<T>::Allocate()
{
    cudaError_t status;
    _pSysStream =   new T*[_subStreams];                      # line 125
    _pDevStream =   new T*[_subStreams];
    _pSysData =     new T[_subStreams * _stride];

    status = cudaMalloc((void **) &_pDevData, _stride * _subStreams * sizeof(T));
    RTERROR(status, (_name+": cudaMalloc in CUDAStream::Allocate failed").c_str());

    for (unsigned int i = 0; i < _subStreams; i++)
    {
        _pSysStream[i] = _pSysData + i * _stride;
        _pDevStream[i] = _pDevData + i * _stride;
    }
}

for which 4.6.3 (Debian 4.6.3-5) happily crunches but 4.7 (Debian 4.7.0-9) gives me

cd /homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/obj-x86_64-linux-gnu/platforms/cuda/sharedTarget && /usr/local/bin/c++   -DOpenMMCuda_EXPORTS -DOPENMM_LIBRARY_NAME=OpenMM -DOPENMM_MAJOR_VERSION=1 -DOPENMM_MINOR_VERSION=0 -DOPENMM_BUILD_VERSION=0 -DOPENMM_SVN_REVISION=\"exported\" -DOPENMM_COPYRIGHT_YEARS=\"2008\" -DOPENMM_AUTHORS=\"Peter.Eastman\" -O3 -DNDEBUG -fPIC -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/cuda/src -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/cuda/./include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/src -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/libraries/validate/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/reference/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/libraries/lbfgs/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/libraries/sfmt/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/libraries/lepton/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/libraries/quern/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/libraries/jama/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/olla/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/openmmapi/include -I/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/./include    -DOPENMMCUDA_BUILDING_SHARED_LIBRARY -o CMakeFiles/OpenMMCuda.dir/__/__/__/src/cuda/kApplyConstraints.cu_OpenMMCuda_generated.cpp.o -c /homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/obj-x86_64-linux-gnu/src/cuda/kApplyConstraints.cu_OpenMMCuda_generated.cpp
/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/cuda/./src/kernels//cudatypes.h: In member function ‘virtual void CUDAStream<T>::Allocate()’:
/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/cuda/./src/kernels//cudatypes.h:125:27: error: capture of non-variable ‘CUDAStream<T>::_subStreams’ 
/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/cuda/./src/kernels//cudatypes.h:68:10: note: ‘unsigned int CUDAStream<T>::_subStreams’ declared here
/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/cuda/./src/kernels//cudatypes.h:126:27: error: capture of non-variable ‘CUDAStream<T>::_subStreams’ 
/homeLvm/moeller/alioth/debichem/experimental/openmm-4.1/platforms/cuda/./src/kernels//cudatypes.h:68:10: note: ‘unsigned int CUDAStream<T>::_subStreams’ declared here
make[3]: *** [platforms/cuda/sharedTarget/CMakeFiles/OpenMMCuda.dir/__/__/__/src/cuda/kApplyConstraints.cu_OpenMMCuda_generated.cpp.o] Error 1


Thanks and regards,

Steffen

N.B. I've changed the Summary, the "non-variable" in the diagnostic refers to a non-automatic variable, in this case a data member, which isn't allowed in a lamdba capture. The problem is not that the error says a member variable is a non-variable, it's that a new-expression is mistaken for a lambda-expresion.


(In reply to comment #2)
> Hello, I have something similar in the OpenMM code
> (https://simtk.org/home/openmm), with no ()s following the "new" as in

Please provide a minimal, complete example, those snippets do not allow your report to be reproduced or confirmed.  This seems to be equivalent but works fine:

template <typename T>
struct CUDAStream
{
    unsigned int    _subStreams;
    T**             _pSysStream;
    void Allocate();
};

template <typename T>
void CUDAStream<T>::Allocate()
{
    _pSysStream =   new T*[_subStreams];
}

Hello, this took me a while.

It seems like the problem is in (or in the interaction with) nvcc, the NVidia compiler. For that you program for the graphics card and the host processor at the same time. The .cu files indicate such files that are not meant to be seen by the regular compiler at a first sight. It then generates preprocessed files for gcc to digest.

$ rm -f t.cu_OpenMMCuda_generated.cpp; /usr/bin/nvcc t.cu -cuda -o t.cu_OpenMMCuda_generated.cpp && g++ -c t.cu_OpenMMCuda_generated.cpp
t.cu: In member function ‘void CUDAStream<T>::Allocate()’:
t.cu:12:27: error: capture of non-variable ‘CUDAStream<T>::_subStreams’ 
t.cu:4:10: note: ‘unsigned int CUDAStream<T>::_subStreams’ declared here

$ cat t.cu
template <typename T>
struct CUDAStream
{
    unsigned int    _subStreams;
    T**             _pSysStream;
    void Allocate();
};

template <typename T>
void CUDAStream<T>::Allocate()
{
    _pSysStream =   new T*[_subStreams];
}


The NVCC-generated code now has the ()s again:

# 9 "t.cu"
template< class T> void
# 10 "t.cu"
CUDAStream< T> ::Allocate()
# 11 "t.cu"
{
# 12 "t.cu"
(_pSysStream) = (new (T *[_subStreams]));
# 13 "t.cu"
}

4.6 crunges it, 4.7 does not. Sounds like we should prepare an email to NVidia. Anybody with any direct contacts?

Thanks

Steffen

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2012 NVIDIA Corporation
Built on Thu_Apr__5_00:24:31_PDT_2012
Cuda compilation tools, release 4.2, V0.2.1221

as shipping with Debian unstable in package version 4.2.9-1.

Anything to do with the cuda compiler is irrelevant here, this report is for a GCC bug not for other programs, even if they generate code that triggers the bug.

Anyway, I'm confirming this, here's a reduced testcase that produces a slightly different eror, due to using a global instead of a data member:

struct A { };

extern int n;

void f()
{
    A** p = new (A*[n]);
}

Accepted by 4.6 and earlier, rejected by 4.7 and later

$ g++-4.7 -c t.cc
t.cc: In function 'void f()':
t.cc:7:21: warning: capture of variable 'n' with non-automatic storage duration [enabled by default]
t.cc:3:5: note: 'int n' declared here


N.B. 4.5 says

t.cc: In function ‘void f()’:
t.cc:7:19: warning: lambda expressions only available with -std=c++0x or -std=gnu++0x

But that was fixed for 4.6

(In reply to comment #6)
> N.B. 4.5 says
> 
> t.cc: In function ‘void f()’:
> t.cc:7:19: warning: lambda expressions only available with -std=c++0x or
> -std=gnu++0x
> 
> But that was fixed for 4.6

I tested 4.5.2, that should be fixed in 4.5.3, see Bug 46159 and its dup Bug 47651

Author: jason
Date: Fri Jun  1 21:19:22 2012
New Revision: 188127

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=188127
Log:
	PR c++/52725
	* parser.c (cp_parser_binary_expression): Bail early if we're parsing
	tentatively and the LHS has a parse error.

Added:
    trunk/gcc/testsuite/g++.dg/parse/new6.C
Modified:
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/parser.c
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/g++.dg/template/sizeof-template-argument.C

Author: jason
Date: Fri Jun  1 21:19:36 2012
New Revision: 188128

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=188128
Log:
	PR c++/52725
	* parser.c (cp_parser_binary_expression): Bail early if we're parsing
	tentatively and the LHS has a parse error.

Added:
    branches/gcc-4_7-branch/gcc/testsuite/g++.dg/parse/new6.C
Modified:
    branches/gcc-4_7-branch/gcc/cp/ChangeLog
    branches/gcc-4_7-branch/gcc/cp/parser.c
    branches/gcc-4_7-branch/gcc/testsuite/ChangeLog
    branches/gcc-4_7-branch/gcc/testsuite/g++.dg/template/sizeof-template-argument.C

Fixed for 4.7.1.