New optimizer produces numerically unstable code

[avl111]

I tried Intel Fortran 10.0.027 for Windows for compiling the EXCITING ab-initio software package (http://exciting.sourceforge.net) The program compiled with /O3 /QxT /Qipo /Qopenmp and Intel MKL works fine on an Intel Core 2 Duo machine. However, when I compile with optimizations for other CPUs: QxN, QxW, QxP, QxO, I get calculation errors from different subroutines, e.g.:

Error(seceqnfv): diagonalisation failed
ZHPGVX returned INFO = 232
The leading minor of the overlap matrix of order 5
is not positive definite
Order of overlap matrix : 227

I tried this on the same Core 2 Duo machine and also on Pentium 4, Pentum 4 HT, and AMD 64x2 with the same negative results. The same code compiled with Fortrtan 9.1 runs fine.

[TimP]

Intel compilers have never supported numerical stability without additional options (-Op for ifort 8, -fp:precise for ifort 9.1). As additional aggressive optimizations are added, it becomes more important for you to use appropriate options, such as -assume:protect_parens -prec-div -prec-sqrt when accuracy is as important as speed. Recent CPUs have reduced the speed advantage of those aggressive optimizations.
If those options don't work in your problem, you should file a problem report on your premier.intel.com account.
I suppose a majority of users would prefer that these options should be the default, but marketers rule here. Until major benchmarks change their rules so as not to penalize compilers which bundle risky optimizations into basic flags, you will get default correct behavior only with open source compilers such as gfortran.
Even with gfortran, you have a problem, as several important optimizations are bundled together with excessively risky ones in the -ffast-math compile flag.