Re: RFA: fix PR rtl-optimization/23898

[Joern RENNECKE <joern dot rennecke at st dot com>]

David Edelsohn wrote:

> 	This patch causes a 10% performance regression to SPECint 254.gap
> on PowerPC.
>
> David
>  
The patch brought the actual behaviour of the compiler in line with what the
comments said it was supposed to do.  But maybe the heuristics are not 
aggressive
enough for optimum powerpc performance.
According to rs6000.md, the maximum branch size is 8, and the minimum is 4.
So we've halved the maximum size of blocks that are copied.

In copy_bb_p, there is this code:

 if (code_may_grow && maybe_hot_bb_p (bb))
   max_size *= 8;

If you change the magic number 8 to 16, do you get the old performance back?
If so, I think we should make this a  value that can be adjusted with 
--param,
and set a suitable rs6000 default.