Inline assembly questions

[Richard Henderson]

On Sun, Apr 06, 2003 at 11:54:53PM -0700, Zack Weinberg wrote:
> Because there is no absolute-jump-to-immediate instruction on the x86.
> The hardware instruction that "call *%cs:addr" maps to is PC-relative.
> You have no choice but to use a scratch register.

Well, there's lcall, but it's *so* much slower that you
don't ever want to use it unless you really are changing
segments at the same time.

> Also, the above constitutes lying to the compiler about control flow;
> the correct thing to do is
> 
> {
>   void (*ptr)(void);
>   asm ("lea %%cs:SomeFunction, %0" : "=r" (ptr))
>   ptr();
> }

Which is also silly because (1) segment overrides are 
ignored for lea and (2) this is equivalent to 

  void (*ptr)(void);
  ptr = &SomeFunction;
  ptr();

which the compiler will happily collapse to

  SomeFunction();

which is what you should have written in the first place.

> > Also, how do I make a function in a GCC C file that is strictly
> > inline assembly be completely bare with no prologue/epilogue (stack
> > frame) or optimizations?
> 
> You write an .s file and feed it directly to the assembler.

Indeed.


r~