Re: [Lightning] Re: lightning and i386

[Shawn W.]

On Aug 12, 2007, at 6:56 AM, Shawn W. wrote:
> Okay. One issue solved, one to go... I've got lightning working on  
> i386, but the functions it's producing aren't returning the same  
> results as the ones on powerpc. Progress of a sort.

And... figured that out. A bug in 1.2c compared to 1.2 (Remember,  
I've been trying to use 1.2c to get amd64 support. The plt version  
someone mentioned won't work for my purposes, even though it has  
amd64 support, because it only works with function calls with up to 3  
arguments. Some of the bytecode I'm trying to turn into native binary  
ends up using more arguments to functions. Unless that's just for how  
many arguments you can pass to lightning-generated functions, not how  
many you can pass to external functions the generated functions are  
calling?)

Anyways, the following snippet:
  f = (pifi)(jit_set_ip(codebuffer).iptr);
  jit_leaf(1);
  arg = jit_arg_i();
  jit_getarg_i(JIT_R0, arg);
  jit_movi_i(JIT_R1, 1);
  jit_eqr_i(JIT_V0, JIT_R0, JIT_R1);
  jit_movr_i(JIT_RET, JIT_V0);
  jit_ret();

compiles to, using 1.2 i386 (Skipping the prologue bits):
mov    %esp,%ebp
mov    0x8(%ebp),%eax
mov    $0x1,%ecx
cmp    %ecx,%eax
mov    $0x0,%ebx
sete   %bl
mov    %ebx,%eax

and, using 1.2c i386:
mov    %esp,%ebp
mov    0x14(%ebp),%eax
mov    $0x1,%ecx
cmp    %ecx,%eax
mov    $0x0,%ebx
sete   %al
mov    %ebx,%eax

The sete line's causing the problem I'm seeing (I don't know if  
that's responsible for all the incorrect results, but returning the  
equality of two numbers is something many paths in my bytecode take.)  
It's also present in 1.2c amd64.

Okay. I can live without amd64 support for a while if i386 works.

However, another problem with just using 1.2 in my project is that  
what I've done so far uses jit_allocai() in places (To allocate a  
pointer, and then pass its address to a function that sets it after  
doing other stuff), and I don't know how to do the equivalent in 1.2.

--
Shawn W.
address@hidden