>> there any particular advantages you can think of to that approach?
>
>I was just thinking of code density, and symmetry with ARM mode to some
>extent. It results in a total 24-bit comment field, and it doesn't need
>any instructions in the calling code to load a register with an immediate
>value. Then again, maybe SWIs aren't very common if they all live inside
>wrapper functions in libc. I'm wondering how the performance would
>compare when you take into account settting up r12 or whatever.
Symmetry with ARM code doesn't cut a great deal of ice as far as I am
concerned; the two ABIs are going to be different anyway and there seems
no point in trying to make them look the same.
I'm also not convinced by your other arguments. In terms of code density I
think SWIs are fairly irrelevant. A typical program doesn't contain any at
all, and even the libc doesn't have all that many - it's just that the few
that do exist are executed rather frequently. In terms of execution time, I
suspect a two-halfword instruction would also lose out. The kernel would have
to load it with two LDRH instructions since the alignment is unknown, then
mask and combine the two parts. I think that is going to be worse than just
loading a constant into the work register, particularly when you take into
account cache effects.
p.
unsubscribe: body of `unsubscribe linux-arm' to [EMAIL PROTECTED]
