64-bit cleanlinest of gfortran

[Scott Robert Ladd]

I'd sent this to you privately earlier, but intended it for the list in general.

On Tue, 14 Dec 2004 10:52:39 -0800, Steve Kargl
<sgk@troutmask.apl.washington.edu> wrote:
> Scott Robert Ladd suggested that gcc is generating bad SSE/SSE2
> code, and SSE/SSE2 is the default on amd64.  I've tried setting
> -mfpmath=387, but I still get several BUS ERROR and failed test.

Actually, you might need to try -mno-sse or -mno-sse2 to stop all
generation of those instructions. I know this solves problems with
some of my code.

An addendum: While my suggestion above does eliminate several LAPACK
failures (as you mentioned in your response to me), it does not
eliminate the BUS error problem. If you can get a small example and
post it as a bug (or send it my way), I can take a look. Try -mno-mmx;
I know AMD64 has some problems there, too.

You might also use the unfairly-maligned -ffast-math; I've found it to
remove some bugs because it changes the code generation logic.

At the moment, I'm trying to solve the SSE/SSE2 bugs, and have made
some headway.

..Scott