Re: Issue when using jit_label() before a jump or branch

[Franz Flasch]

Hi Paul!

Thanks for your answer! Meanwhile I came to the conclusion, that it might be an internal optimization of GNU Lightning causing this. Considering my code example it does not make any

difference to jump back to the label before the jump. I thought that I had also cases where the optimization was wrong, but at this point I am not sure anymore ':-)

I also tried jit_patch() and jit_indirect() but it does not make any difference at all.

Anyway, thanks for the tips! I didn't know that I can use jit_patch() directly for forward jumps. That's way easier to use.

Thanks, regards,
Franz

On 1/29/22 6:42 PM, Paul Cercueil wrote:

> Hi Franz,
>       
>       
>       I'm not really sure what could be the problem, sorry.
>       
>       
>       One thing I noticed though, is that you use jit_forward(). You
>       actually do not need to use that.
>       
>       
>       Try this:
>       
>       
>       label = jit_label();
>       
>       jump = jit_jmpi();
>       
>       
>       jit_addi(JIT_R0, JIT_R0, 42);
>       
>       
>       jit_patch(jump);
>       
>       
>       jit_addi(JIT_R0, JIT_R0, 42);
>       
>       
>       jump2 = jit_jmpi();
>       
>       jit_patch_at(jump2, label);
>       
>       
>       jit_retr(JIT_R0);
>       
>       
>       A second thing to try: replace your jit_label() with a
>       jit_indirect(). I actually always do that for backwards branches.
>       
>       
>       Cheers,
>       
>       -Paul
>       
>       
>       
>       Le dim., janv. 23 2022 at 13:00:13 +0100, Franz Flasch
>       <franz.flasch@gmx.at> a écrit :
>       
> > Hi!
> >         
> >         
> >         I think there is kind of an issue in GNU lightning when using a
> >         label as jump target if this label was created immediately
> >         before a jump or branch. Here is an example:
> >         
> >         
> >         
> >             /* get a label here */
> >         
> >              label = jit_label();
> >         
> >         
> >              /* Prepare forward jump */
> >         
> >              jump = jit_jmpi();
> >         
> >              forward = jit_forward();
> >         
> >              jit_patch_at(jump, forward);
> >         
> >         
> >              /* Some random instructions */
> >         
> >              jit_addi(JIT_R0, JIT_R0, 42);
> >         
> >         
> >              /* Link forward jump */
> >         
> >              jit_link(forward);
> >         
> >         
> >              /* Some random instructions */
> >         
> >              jit_addi(JIT_R0, JIT_R0, 42);
> >         
> >         
> >              /* Now prepare backward jump to first label */
> >         
> >              jump2 = jit_jmpi();
> >         
> >              jit_patch_at(jump2, label);
> >         
> >         
> >              jit_retr(JIT_R0);
> >         
> >         
> >         
> >         The above instructions result into the following GNU lightning
> >         instructions:
> >         
> >         
> >             L0: %rax /* prolog */
> >         
> >                  jmpi L2
> >         
> >              L4: %rax
> >         
> >                  addi %rax %rax 0x2a
> >         
> >              L2: %rax
> >         
> >                  addi %rax %rax 0x2a
> >         
> >                  jmpi L2
> >         
> >              L5: %rax
> >         
> >                  retr %rax
> >         
> >                   \__ live %rax
> >         
> >                   \__ ret
> >         
> >              L3: /* epilog */
> >         
> >         
> >         
> >         As you can see the second jump is wrong. It jumps back to L2
> >         even though it should jump to L0 (before the first jump). When I
> >         use the label from the first jump like this
> >         
> >         
> >             jit_patch_at(jump2, jump);
> >         
> >         
> >         it works as expected.
> >         
> >         
> >         But why can't I use a normal label here? Is this the intended
> >         behavior or is this a bug?
> >         
> >         
> >         If there is another instruction between jit_label() and
> >         jit_jmpi() it also works as expected.
> >         
> >         
> >         
> >         The same behavior also occurs with branches.
> >         
> >         
> >         
> >         I could just use the first jump label to solve this issue,
> >         however my code generation engine currently does not know about
> >         this in advance so I rely on normal labels to work correctly.
> >         
> >         
> >         
> >         Thanks,
> >         
> >          Franz
> >         
> >         
> >         
> >       
>       
>       
>