-finline-limit=50000 causing issues in parscalarvec branch with gcc-4.6.3

[bjoo]

I've compiled parscalarvec with -finline-limit=50000 and I get an odd result
for a test of

SpinMatrix g_one = 1;

LatticeFermion chi = zero;
LatticeFermion chi2 = zero;
LatticeFermion tmp2 = zero;
gaussian(tmp2);
LatticeFermion tmp3 = g_one*tmp2;

chi = shift(g_one*tmp2 , BACKWARD, j_decay);
chi2 =shift( tmp3, BACKWARD, j_decay);

QDPIO::cout << "Test 29: norm2 ( chi2 - chi )=" << norm2( chi2 -chi ) << endl;

The remaining C++ flags apart from -finline-limit are:

-fopenmp -g -O3 -std=c++0x -fargument-noalias-global 

and the compiler used is

/dist/gcc-4.6.3/bin/g++

The answer ought to be zero.

With -finline-limit=50000 in the parscalarvec branch I get:

Test 29: norm2 ( chi2 - chi )=874836.987165243

Without the -finline-limit option I get:

Test 29: norm2 ( chi2 - chi )=0

as I should, Please beware in your own builds.

PS: I also tried a variety of options here, using for example -O2 -O0, and leaving off the -fargument-noalias-global option. The -finline-limit is the smoking gun.
With the -finline-limit option I also replaced g_one with other types

OLattice<PScalar<PColorMatrix< RComplex< ILattice< REAL64, INNER_LEN> >, 3> > > g_one;
gaussian(g_one);

and

OScalar<PScalar<PColorMatrix< RComplex< IScalar< REAL64 > >, 3> > > g_one = 1;

etc, and all seem to work ok without the -finline-limit=50000 option.