Re: PA problem

[Richard Earnshaw <rearnsha at arm dot com>]

> 
>  In message <200102281706.RAA05092@cam-mail2.cambridge.arm.com>you write:
>   > Of course, the other alternative is to disable the splitter until after 
>   > reload has completed.  I'm not sure which will lead to the best code.
> But we want to schedule those instructions independently.
> 
> Basically what we've done is take something like this
> 
> t1 = a * b;
> t2 = t1 + c
> 
> And turns it into
> 
> t1 = a * b;
> t2 = a * b + c
> 
> Note how t2 is no longer dependent on t1.  That allows the instructions to
> execute and retire independently of each other which gives the scheduler
> more freedom in their placement in the insn stream.
> 

Sure, I realise that (it says so in the comments IIRC).  But there are two 
scheduling passes, one of which is after reload, and hence after the 
pattern can be safely split.