Re: CVS Build problem (what.m missing?)

[John W. Eaton]

On 31-Oct-2007, John Swensen wrote:

| I also have this problem on OSX using the CVS as of 5:18PM EST 
| 10/31/2007.  My configure line is as follows:
| 
| ./configure CC=gcc-4 CXX=g++-4 F77=gfortran CPPFLAGS=-I/sw/include 
| LDFLAGS=-L/sw/lib --prefix=/sw/opt/octave/cvs_2.9_20071007
| 
| The config.log file indicates it failed at:
| configure:15915: checking BLAS library calling convention compatibility
| configure:15958: result: no
| 
| Reading through your email chain, I wasn't sure whether a solution had 
| been reached, other than to cut out the test code so it always succeeds, 
| but is this going to cause problems when compiling/running?
| 
| I tried to compile the conftest code by hand, but if I compile/link it with:
| gfortran conftest.f -c -g3
| gfortran conftest.o -o conftest  -Wl,-framework -Wl,vecLib
| and then try to run it, I get a "Illegal instruction" error.  I ran it 
| in gdb and get further information concerning where the error occurred:
| Program received signal EXC_BAD_INSTRUCTION, Illegal instruction/operand.
| 0x8fe12f94 in __dyld_stub_binding_helper_interface ()

If you step through the program, is the crash happening when zdotu is
called?  If so, then I think you are hitting precisely the problem
that the test intends to protect against, which is to prevent problems
like this from happening when you do something like

  x = 1:10;  y = i*x';  x * y

in Octave.  If you do bypass the test, does Octave crash for this test
case, or do you get the correct answer (0 + 385i), or something else?

Since this problem may only affect a single call to zdotu in Octave, I
suppose another option would be to replace the call to the BLAS
routine in this case with a simple loop (we had that at one point
anyway).

jwe