Re: Bytecode hacks

[Roberto Ierusalimschy <roberto@...>]

> On 5/21/2014 8:27 PM, Luiz Henrique de Figueiredo wrote:
> >>I think there are patchs around for the t:[var]() syntax.
> >Actually, this one seems simple. A quick hack is to just change
> >suffixedexp in lparser.c (from Lua 5.2.3):
> >
> >        case ':': {  /* `:' NAME funcargs */
> >          expdesc key;
> >          luaX_next(ls);
> >-        checkname(ls, &key);
> >+        if (ls->t.token=='[') {
> >+           luaK_exp2anyregup(fs, v);
> >+           yindex(ls, &key);
> >+        }
> >+        else
> >+           checkname(ls, &key);
> >          luaK_self(fs, v, &key);
> >          funcargs(ls, v, line);
> >          break;
> > 	[...]

(Thanks for the debugging!)


> The luac listing of these cases are all of this form:
>     22    [10]    CALL         3 0 2
>     23    [10]    SELF         4 0 3
>     24    [10]    LOADK        5 -10    ; 3
>     25    [10]    LOADK        6 -11    ; 32
>     26    [10]    LOADK        7 -12    ; 33
>     27    [10]    CALL         4 4 1
>
> The C argument of SELF is always a register, the same register as
> the the A argument of the previous CALL.
> Correct bytecode for the case above  should be like this:
>     22    [10]    CALL         3 0 2
>     23    [10]    SELF         4 0 3
>     24    [10]    LOADK        6 -10    ; 3
>     25    [10]    LOADK        7 -11    ; 32
>     26    [10]    LOADK        8 -12    ; 33
>     27    [10]    CALL         4 5 1

The problem is related to 'luaK_reserveregs' (done in 'luaK_self'). The
opcode OP_SELF needs ra + 1 as a temporary, so nothing can go there.  In
the correct code, register 5 (the next after 'ra' in SELF) is left free,
so that SELF can use it.

To fix this bug, I think the patch cannot call 'yindex'. Instead, it
must do what 'yindex' does but without calling 'luaK_exp2val'. (This
is a wild guess ;)

-- Roberto