[LLVMdev] Indirect branch instruction

[Nicolas Geoffray]

Chris Lattner wrote:
> I don't understand.  You're making a claim that the C compiler isn't 
> producing optimal code for some case.  Can you give an example of a C 
> function that llvm compiles to something suboptimal?  LLVM certainly does 
> generate stuff like "jmp %eax", but presumably not in the way you want. 
> Without an example to see what you mean, we won't be able to fix the 
> deficiency that you've encountered.
>
>   
Sorry, I've been misunderstood. I'm claiming nothing. I'm just
assuming that when using a SwitchInst, the just in time compiler
generates two jumps, one to jump to the right case, and then one
to jump to the correct label.

Typically, I understood that, to make indirect branches, the
(pseudo-) code generated is:


label1:
    instructions
label2:
    instructions
label3:
    instructions

my_value = @ of a label
switch(my_value)
{
    case label1: jump label1
    case label2: jump label2
    case label3: jump label3
    default:   abort("invalid label")
}

What I was asking is : what is the assembly code generated by llvm for this
kind of code? My guess was 2 jumps. But it might be 1 jump, based on
the fact that llvm knows it switches on labels.

I hope it's clearer now.

Thanks Chris,

Nicolas
> Thanks,
>
> -Chris
>
>