Re: Two entry points for a function

[Marc Nieper-Wißkirchen]

Am Mo., 8. Aug. 2022 um 23:16 Uhr schrieb Paul Cercueil <paul@crapouillou.net>:

Le lun., août 8 2022 at 23:04:02 +0200, Marc Nieper-Wißkirchen
<marc.nieper+gnu@gmail.com> a écrit :
> Hi Paul,
>
> thanks for chiming in.
>
> I didn't know about jit_align.  Maybe it should be documented (along
> with some instruction that embeds given bytes in the code stream?).
>
> So what I have in mind could probably be done with GNU lightning as
> follows:

I guess that works, yes. Be careful that your "short piece of code"
cannot do much (Lightning won't save/restore the callee-saved registers
for you, stack won't work, jit_ret() won't work, etc).

I have been assuming that I am jumping to entry1/entry2 from a function that has a frame set up as in

jit_prolog ();

jit_frame (256);

...

jit_movi (JIT_R0, (jit_word_t) addr1 + 16);

jit_jmpr (JIT_R0);

-Paul

> jit_prolog ();
> jit_tramp (256);
> jit_align (16);
> entry1 = label_indirect ();
> /* short piece of code */
> jit_align (16);
> entry2 = label_indirect ();
> ...
> jit_epilog ();
> jit_emit ();
> char *addr1 = jit_address (entry1);
> char *addr2 = jit_address (entry2);
> assert (addr2 - addr1 == 16);
> ...
>
> Am Mo., 8. Aug. 2022 um 22:53 Uhr schrieb Paul Cercueil
> <paul@crapouillou.net>:
>> Hi Marc,
>>
>>  I don't think you can have multiple entry points of a function (or
>>  maybe with jit_indirect()?) but you can cheese things a bit with a
>>  trampoline.
>>
>>  Here's what I do, more or less:
>>
>>  ----
>>
>>  jit_prolog();
>>  jit_tramp(256);
>>
>>  for (i = 0; i < nb_entry_points; i++) {
>>      jit_addi(JIT_R0, 1);
>>      jit_align(8);
>>  }
>>
>>  jit_epilog();
>>  jit_prolog();
>>
>>  /* Read JIT_R0, its value give us what entry point was taken */
>>
>>  jit_epilog();
>>
>>  ----
>>
>>  In that example my JIT_R0 is always 0 on entry. The jit_align()
>> makes
>>  sure that each one of your entry points is aligned to 8 bytes.
>> There is
>>  still some decoding needed (I read a C function pointer from a
>> table at
>>  the index pointed by JIT_R0), but the decoding is done in that
>> function
>>  and not in the callers.
>>
>>  Cheers,
>>  -Paul
>>
>>  Le lun., août 8 2022 at 22:16:28 +0200, Marc Nieper-Wißkirchen
>>  <marc.nieper+gnu@gmail.com> a écrit :
>>  > I am looking for a way to implement a function (pointer) with GNU
>>  > lightning that has two entry points.
>>  >
>>  > In native assembly, I would code something like this:
>>  >
>>  >         .align 16
>>  > entry1: jmp L0
>>  >         .align 16
>>  > entry2: ...
>>  >         ...
>>  > L0:     ...
>>  >
>>  > The function would be represented by a pointer to entry1.  If I
>>  > wanted to jump to the second entry point, I would just add 16 to
>> the
>>  > pointer.
>>  >
>>  > How can a similar thing be achieved with GNU lightning?  So far,
>> the
>>  > best that has come to my mind is to use one more indirection.  The
>>  > "function pointer" points to a structure having the addresses of
>>  > entry1 and entry2 as fields.  (Let us assume we are in a
>> trampoline
>>  > so that the stack is already set up).  But this is less efficient
>> due
>>  > to the extra indirection.
>>  >
>>  > Is there a way to emulate the above native code with GNU
>> lightning?
>>  > And, if not, what would have to be added?
>>  >
>>  > Marc
>>
>>