On Fri Nov 10 17:36:05 2006, mdiep wrote:
> This was taken from t/pmc/iterator.t:
>
> # XXX
> # swapping the next two lines breaks JIT/i386
> # the reason is the if/unless optimization: When the
> # previous opcode sets flags, these are used - but
> # there is no check, that the same register is used in the "if".
> inc I0
> dec I1
> if I1, fill
>
> I've taken this comment out of the test file because it's in no way
> relevant to Iterator. It needs to be confirmed or rejected. And if it
> *is* confirmed, it needs to be added as a test somewhere else.
>
Well, in core.jit I see:
TEMPLATE Parrot_ifunless_x_ic {
/*
* FIXME - the compare prev_op[1] == cur_op[1] should do it
* dec I1
* inc I0
* if I1, boom
*/
if ( P_ARITH && jit_info->prev_op[1] == jit_info->cur_op[1]) {
It's not clear to me from that comment if the text after FIXME explains
that the test does not work or that it was added to do the fix. I
suspect the second - it's the Right Thing to do as far as I can see, and
trying to re-create the bug does not show it up. I swapped those lines
round on the test, and it still passes under -j.
I think maybe this bug is resolved and the FIXME comment should go, but
a comment from someone who's worked on the JIT would be nice.
Jonathan
