Isel opcode unrecognized opcode?

tying to compile a small test.

using

GCC -mpowerpc -misel test.c

GCC knows -misel, but if I add ISEL to inline asm code, it does not work.

Also tried.

GCC -misel=yes test.c

but -misel=yes is reported as none supported, so -misel should work, I think.


FSEL works, but there are some issues, you can't move from GPR to FR, whit out storing into memory and loading from memory.

and there is no bit operations like "AND" oppcode for float registers, so it looks like I have do part of what I'm trying to do whit integer and the other part whit float, that sucks.

@LiveForIt

Looking at the GCC doc online, I see:

Quote:

So this option is about generation of the instruction. You certainly have to choose a CPU that supports the instruction.

Note that currently, we have no platform that supports this instruction.

To compile with the inline isel instruction, try something like:
gcc -mcpu=e500mc -o isel isel.c
or on recent compilers:
gcc -mtune=e500mc -o isel isel.c

Note that if you do that on AmigaOS, that will crash with an undefined instruction on our machines.

And to target the e5500 core, if the compiler does not support it, you will tune for e500mc, what could be worse than not use isel instruction.

But that's always very good to experiment !

@corto

Thanks, but I have also problem converting word into float.

any idea?

As you said in your first post, you can't move directly a GPR into a FPR. The integer value must be saved in memory and then loaded into a FPR with instruction lfiwax (Load Floating-Point as Integer Word Algebraic Indexed).

For the reversed operation, a short C test with the generated asm code shew that there is a conversion:
fctiwz 0,30
before saving the register:
stfiwx 0,0,9
But this part should be investigated more. I am not very comfortable with FP instruction set.

@corto

http://wiird.l0nk.org/forum/index.php ... on=printpage;topic=6593.0

save value into int var

i = 10;



cast int var into float var.

f = (float) i;



I think we can say that casting between float and int is sloooow.