[Bug tree-optimization/16962] loop unrolling with x86-64 asm not efficient

["tomstdenis at iahu dot ca" <gcc-bugzilla@gcc.gnu.org>]

------- Additional Comments From tomstdenis at iahu dot ca  2004-08-10 14:13 -------
(In reply to comment #1) 
> Please use the attachment function for large files in the future. Note also 
> we don't want assembly output anyway (as stated on 
http://gcc.gnu.org/bugs.html). 
 
Ooops sorry. 
 
> Loop work is currently only being done at the lno-branch. It would be nice 
> if you could test it there. 
 
I'll have to ask.  I'm using someone else box atm. 
  
> Also, this way of unrolling loops doesn't seem fundamentally wrong to me 
> Can you provide performance numbers that show that it is worse? 
 
No because I can't get it to compile the other way ;-( that's the point! 
 
>From what I can see the bug [maybe in my code]?  Is that I do things like  
 
mulq (%1) 
 
So GCC doesn't realize it can mod that and do  
 
mulq 0(%1) 
...next iteration 
mulq 8(%1) 
...next iteration 
mulq 16(%1) 
 
So instead it does 
 
mulq 0(%1) 
... 
lea 8(%some_register),%some_other 
mulq (%some_other) 
... 
 
Is my ASM code just wrong (in that I mean I'm not making best use of it?) or 
is this a legit chance for GCC to optimize better? 
 
>  
 
 

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
  GCC build triplet|x86_64-linux                |x86_86-linux
   GCC host triplet|x86_64-linux                |x86_86-linux
 GCC target triplet|x86_64-linux                |x86_86-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16962