Re: [ira] patch for optimistic register coalescing implementation

[Vladimir Makarov <vmakarov at redhat dot com>]

Seongbae Park wrote:
> I very quickly skimmed through the patch, so I could have easily missed
> but I couldn't tell where undo coalescing is handled,
> which makes the optimistic coalescing different from simple aggressive
> coalescing
> (and more expensive in compile time too).
>   
It is done at the end of ira-color.c::assign_hard_reg where there is 
printf with  "Pushing ".  If the coalesced allocnos failed to get a 
hard-registers, they all are uncoalesced and put onto the stack (regular 
coalescing based on hard register and conflicting costs works after 
that).  There are few things to investigate:

o Different strategies of splitting (not all at once) in the failure 
case. I don't think it can give visible improvement because coalescing 
based on register costs works well.  Also other strategies will slow 
down RA (may be significantly in some cases).

o Removing the pseudo-register moves in RA instead of assigning the same 
hard register to the coalesced allocnos which usually results in trivial 
removing the same hard register move in the post-reload pass.  I don't 
think removing the pseudo-register moves in RA will be more profitable.  
With saving the moves in RA if the reload decides to spill a 
pseudo-register, pseudo-registers coalesced with given pseudo-register 
will still stay in the hard-register (which usually is more profitable 
than spilling all coalesced pseudo-registers).
> Can you point me to which part of the code handles that ?
> Apology if I missed the obvious.
>