Re: Lightning register liveliness

[Paulo César Pereira de Andrade]

Em qua., 22 de jan. de 2020 às 15:23, Paulo César Pereira de Andrade
<address@hidden> escreveu:
>
> Em qui., 9 de jan. de 2020 às 00:26, Paul Cercueil
> <address@hidden> escreveu:
> >
> > Hi Paulo,
>
>   Hi Paul,
>
>   Sorry for the delay. Just found this message I previously missed.
>
> > I am facing a problem related to register liveliness. I could only
> > reproduce it on MinGW, I never faced it on Linux x86_64 or MIPS.
> >
> > Here is an example code to reproduce the issue (sorry for the length).
> > It is not important to understand what the code does; what is important
> > to see, is that my JIT_V0 (== rbx), which is set at the very beginning
> > of the code and read back at the last opcode (provided that the node2
> > beqi branch is taken), is overwritten by the call to jit_ltr_u.
> >
> > I know that Lightning assumes that the caller-saved registers are dead
> > after a function call, but there's no function call here.
> >
> > Could you give me some light in why Lightning uses my registers? Do I
> > have to use jit_live() somewhere here?
>
>   I understand what is happening. It is a bug in the logic of live registers,
> that does not mark JIT_V0 live, and then it is used in the ltr_u translation,
> when it needs a special kind of register. Due to the way it handles the live
> state, it thinks it is dead in 'jit_movi(JIT_V0, 0x80065ff8);' and does not
> properly follow the previous branch in 'node2 = jit_beqi(JIT_V1, 0);'.
>   I will work on a patch, and let you know about it when it is ready.
>
>   An ugly hack, for a temporary solution would be to add a jit_live as below:
>
>    jit_subi(CYCLE_REG, CYCLE_REG, 0x14);
> +  jit_live(JIT_V0);
>   node2 = jit_beqi(JIT_V1, 0);

  A patch was now pushed. This code should be revisited, but
the added patch should correct any variant of the problem you
experienced.

> > Thanks,
> > -Paul
> >
>
> Thanks,
> Paulo