llvm dev - Feb 2012 - [LLVMdev] Remove redundant code after frame index elimination

Hi list,

I added custom code to eliminate frame index references. I replace each 
FI reference with a subtraction from my frame pointer register (just 
like ebp in x86). Its result is stored in another register which is used 
by the load/store instructions. Nevertheless, this operation gives me 
redundant subs that I would like to remove after all fi references have 
been eliminated. For example:

r = load fi<1> -->
{
   sp = subm fi, cst_offset
   r = load sp
}
...
store fi<1>, val -->
{
   sp = subm fi, cst_offset
   store sp, val
}

Any advice on what is the better solution to remove that code ? Should I 
create a custom pass after the prolog/epilog insertion one ?

Ivan

Hello again,

I checked the sparc BE and I've seen that it adds also some MI when it 
eliminates FI references. Nevertheless, the testcase is very rare and it 
hardly happens in real applications but it can describe my problem.

void test() {
   short a1;
   ...
   short a6003;

   a6000 = 2;
   a6000 = 3;
}

     sethi 4194292, %g1
     or %g1, 184, %g1
     save %sp, %g1, %sp
     or %g0, 2, %l0
     sethi 4194292, %g1
     add %g1, %fp, %g1
     sth %l0, [%g1+288]
     or %g0, 3, %l0
     sethi 4194292, %g1
     add %g1, %fp, %g1
     sth %l0, [%g1+288]
     jmp %i7+8
     restore %g0, %g0, %g0

Before the 2nd STH instruction, redundant code is added by the FI 
elimination routine (%g1 was already defined with the same value). My 
target does not have an addressing mode to represent fp + offset, so I 
need to store the result in an intermediate register to access the memory.

It seems to me that the FI abstract model implies an indexed absolute 
addressing support provided by the target.
Do I need to rely in early optimizations phases to avoid this kind of 
cases ? or, Is there a way to lower the FI node to extract its offset 
and write an explicit add operation ?
I found that at the target lowering phase, no FI offset is represented 
for different variables (it's just a slot). It's defined at the FI 
elimination phase though. I could not find a way to lower it because of 
that.

Please, any help would be appreciated.

Regards,
Ivan



Le 10/02/2012 13:47, Ivan Llopard a écrit :
> Hi list,
>
> I added custom code to eliminate frame index references. I replace 
> each FI reference with a subtraction from my frame pointer register 
> (just like ebp in x86). Its result is stored in another register which 
> is used by the load/store instructions. Nevertheless, this operation 
> gives me redundant subs that I would like to remove after all fi 
> references have been eliminated. For example:
>
> r = load fi<1> -->
> {
>   sp = subm fi, cst_offset
>   r = load sp
> }
> ...
> store fi<1>, val -->
> {
>   sp = subm fi, cst_offset
>   store sp, val
> }
>
> Any advice on what is the better solution to remove that code ? Should 
> I create a custom pass after the prolog/epilog insertion one ?
>
> Ivan