This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: LAPACK: good news and bad news



[ adding gcc@gcc.gnu.org as this might be of general interest; for those who are not in the know, the DOUBLE COMPLEX version of BLAS used to fail its tests until not long ago, the new failure Steve discusses below is now PR 15982 ]


Steve Kargl wrote:
The good news is that someone committed a fix to gfortran
that now permits the BLAS included with LAPACK to pass all
of its test program.

This is most likely due to 2004-06-10 Roger Sayle <roger@eyesopen.com>

	* builtins.c (expand_builtin): Fall back to library function call for
	conj, conjf, conjl, creal, crealf, creall, cimag, cimagf and cimagl.
	(fold_builtin_1): Lower built-ins BUILT_IN_CONJ{,F,L} to CONJ_EXPR,
	BUILT_IN_CREAL{,F,L} to REALPART_EXPR, and BUILT_IN_CIMAG{,F,L} to
	IMAGPART_EXPR respectively.

No other patch during the last few days seems to have touched complex arithmetic, and the bug was almost certainly not in the frontend, as we treat COMPLEX*4 and COMPLEX*8 completely analogously AFAIK. Tomorrow, I'll try to distill a smaller-than-BLAS testcase with this information, which fails with an older gfortran and now passes, so that we don't regress again.

The bad news is that LAPACK no longer
compiles. I'm getting

Too bad, but we're getting closer.


- Tobi


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]