how to test library generation

[Thomas Koenig]

On Sun, Aug 28, 2005 at 07:27:45AM -0700, Tim Prince wrote:
> I haven't yet understood how modifications in the .m4 files are 
> propagated, and how they should be tested.

What I do is pretty horrible, but at least it works, and I haven't
found anything else that works.

I copy the libgfortran directory to another location, then run
./configure --enable-maintainer-mode in that directory.  This
fails with an error, but only after the Makefile has been generated.

If I then change the m4 files and type 'make generated/*.c', the
files are correctly regenerated.  I then only need to copy over
the m4 and generated files over to the original directory.

> I was concerned about propagating excessive source code unrolling into 
> the complex variants; is that a problem?

You mean because of register pressure?

> Much of the library will not build with gcc-4.1, except at default 
> options.    I don't see much point in performance tests, without trying 
> normal options, such as
> -march=nocona -O2 -funroll-loops
> which provokes repeated gcc 4.1 internal errors (not in the part I have 
> been modifying).

This is bad.  Seems like a few PRs are needed...

> gcc-4.0.2 is OK, but the library source which goes with that is 
> different,

Not too much.  The difference is mostly cleanups to avoid warnings.

>and I think 4.0.2 misses optimization found by 4.1

Probably.