[Tarantool-patches] [PATCH luajit] Fix predict_next() in parser.

[Sergey Bronnikov]

Hi, Sergey


thanks for the patch! See comments inline.


On 8/15/23 17:25, Sergey Kaplun wrote:
> From: Mike Pall <mike>
>
> Reported by Sergey Kaplun.
>
> (cherry-picked from commit caf7cbc57c945f7b68871ad72abafb2b6e6fb7f5)
>
> Assume, we have the following Lua code:
> | local _
> | for _ in (nil):foo() do end
>
> The first part of the bytecode emitted for it is the following:
> | 0001    KNIL     0   1
> | 0002    MOV      2   1
> | 0003    TGETS    1   1   0  ; "foo"
> | 0004    CALL     1   4   2
>
> The `0001 KNIL` is a result of merging two `KPRI` instructions: one for
> the local variable, one for the slot with `nil` object. During parsing in
> `predict_next()` the second `MOV` bytecode is examined to set `pairs` or
> `next` local variable. But, as far as it moves `nil` value, that isn't
> an actual variable, so it has no the name this leads to the crash.
>
> This patch adds the check to be sure that `RD` in the `MOV` bytecode is
> an actual variable.
>
> Sergey Kaplun:
> * added the description and the test for the problem
>
> Part of tarantool/tarantool#8825
> ---
>
> Branch: https://github.com/tarantool/luajit/tree/skaplun/lj-1033-fix-parsing-predict-next
> PR: https://github.com/tarantool/tarantool/pull/8987
> Related issues:
> * https://github.com/LuaJIT/LuaJIT/issues/1033
> * https://github.com/tarantool/tarantool/issues/8825
>
>   src/lj_parse.c                                |  1 +
>   .../lj-1033-fix-parsing-predict-next.test.lua | 30 +++++++++++++++++++
>   2 files changed, 31 insertions(+)
>   create mode 100644 test/tarantool-tests/lj-1033-fix-parsing-predict-next.test.lua
>
> diff --git a/src/lj_parse.c b/src/lj_parse.c
> index 3f6caaec..420b95cb 100644
> --- a/src/lj_parse.c
> +++ b/src/lj_parse.c
> @@ -2532,6 +2532,7 @@ static int predict_next(LexState *ls, FuncState *fs, BCPos pc)
>     cTValue *o;
>     switch (bc_op(ins)) {
>     case BC_MOV:
> +    if (bc_d(ins) >= fs->nactvar) return 0;
>       name = gco2str(gcref(var_get(ls, fs, bc_d(ins)).name));
>       break;
>     case BC_UGET:
> diff --git a/test/tarantool-tests/lj-1033-fix-parsing-predict-next.test.lua b/test/tarantool-tests/lj-1033-fix-parsing-predict-next.test.lua
> new file mode 100644
> index 00000000..624344eb
> --- /dev/null
> +++ b/test/tarantool-tests/lj-1033-fix-parsing-predict-next.test.lua
> @@ -0,0 +1,30 @@
> +local tap = require('tap')
> +local test = tap.test('lj-1033-fix-parsing-predict-next')
> +
> +test:plan(3)
> +
> +local res_f = loadstring([[
> +-- This local variable is necessary, because it emits `KPRI`
> +-- bytecode, with which the next `KPRI` bytecode will be merged.
> +--
> +-- The resulting bytecode is the following:
> +--
> +-- 0001    KNIL     0   1
> +-- 0002    MOV      2   1
> +-- 0003    TGETS    1   1   0  ; "foo"
> +-- 0004    CALL     1   4   2
> +--
> +-- This MOV don't use any variable value from the stack, so the
> +-- attempt to get the name in `predict_next() leads to the crash.
What is a point to put a comment inside loadstring, not before it?
> +local _
> +for _ in (nil):foo() do end
> +]])
> +
> +test:ok(res_f, 'chunk loaded sucsessfully')
typo: sucsessfully -> successfully
> +
> +local res, err = pcall(res_f)
> +
> +test:ok(not res, 'loaded function not executed')

it is not clear for me what for do you need checking result code. I 
would omit it.

Feel free to ignore.


> +test:like(err, 'attempt to index a nil value', 'correct error message')
> +
> +test:done(true)