Re: [PATCH] Improved target tuning in simplify-rtx.c

[Jim Wilson <wilson at specifixinc dot com>]

Roger Sayle wrote:
> quality of RTX_COSTS improved.  It might come as a shock to some that
> the rs6000 backend doesn't provide timing for floating point addition
> and multiplication 

Defining RTX_COSTS for FP instructions traditionally didn't do anything 
useful.  If you look at where rtx_costs is used, it is for things like 
comparing the costs of two equivalent sequences.  This is common in 
integer code, where you can use a shift/add sequence or a multiply for 
instance.  But this is not true for FP code.  If you need an FP add, 
then you emit an FP add.  There is no other alternative, and hence no 
need for cost information.

Looking at all of the rtx_cost calls, I see only one that looks vaguely 
useful for FP instructions.  That is the call in loop-invariant.c.  This 
is a new code added after the gcc-3.4 branch.  So that implies that 
defining costs for for FP instructions may be useful post gcc-3.4, but 
not before then.

We do need timing info for scheduling.  This is very important for FP 
instructions.  However, the scheduler does not use rtx_cost, and it 
never has.  The scheduler uses either the new DFA scheduler stuff in the 
md file, or the old style define_function_unit stuff.  This information 
is present in the rs6000 port.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com