3DNow in gcc 3.2

[Klaus Niederkrueger <kniederk at MI dot Uni-Koeln dot DE>]

Hi,

Yesterday I played a bit with the gcc 3.2 on an AMD k6-2 to test the
__builtin_ia32_XXX() functions and I am very, very happy.

But there are still some doubts: First I have to say that I did not
install the gcc-3.2, but I left it in my home directory and I compiled
things with

~/gcc-3.2/gcc/xgcc -B ~/gcc-3.2/gcc/ ....

so that maybe this is what produces the errors, I am seeing.

O.k. then. I made a program, where I have a floating point array and I do
some 3DNow-computations to fill the array with numbers. After doing that I
print the array out. All values are correct, but the first one, which is
printed as "nan". The strange thing is that I debug the program, and if I
use "print arr" in the debugger, then the whole array is correct -- also
the first element.

What am I doing wrong?

Also I experienced a lot of crashes of the GCC, when I used
"__builtin_ia32_femms()" (in fact I was not able to use this instruction
at all without making the compiler crash).

Finally I would like to ask, what is meant in the manual of the gcc, by
saying that you can add, substract etc. vector type variables. Does this
mean that the compiler chooses the best "__builtin_ia32_XXX"-instruction
for each case to do what I am asking for or does this mean that the
compiler makes up normal floating point code to do add the
vector-variables?

Thanks for your answers

	Klaus