Re: Higher level RTL issues

[Daniel Berlin <dan at cgsoftware dot com>]

On Monday, October 22, 2001, at 05:04  PM, law@redhat.com wrote:

>   In message <BF0CB0D7-C728-11D5-9F25-0030657B5340@cgsoftware.com>you 
> write:
>>> So how is this done when SSA names are not reflected in the IR?
>> You can do the same thing.
>> You just need names.
>
>> If you look at the ssa-ccp patch i submitted, you'll note i added a
>> unique id number to each ref, and use it in the following way to give 
>> an
>> ssa name:
>> For a phi or a def, the ssa name is the id number.
>> For a use, the ssa name is the id number of the associated def.
>> This gives you the same "name" an explicit renamed representation would
>> give you, but we still haven't rewritten any code.
> Right, but whenever you make this kind of transformation you have to 
> scrurry
> around and find all the uses the update the id number to point to the 
> new
> def.
None that you wouldn't have to do anyway, since you need to update the 
use's chain to point to the new def anyway.  The ID number is part of 
the def/use structure, so doing the required update makes it work.
Though i think this is what you meant, i'm just having a bit of trouble 
parsing your sentence. Criminal law class does it to me.
Remember, we have all the uses for a def, too, so when you remove the 
def, you can just automatically update all the uses to the new def 
pretty simply (since, in the case of dominator optimizations, we know 
which def is the new one).



> Ugh.  How unpleasant.  The beauty of a rewriting SSA is it just works.
>
It's not actually as unpleasant as one might think, because all the 
updating can be done in replace_expr_in_tree, without *too* much trouble.
So from a programming perspective, it's quite possible to make it so it 
is never seen by the ssa passes.


>> I'm assuming you keep the definitions/uses/phi links updated
>> approriately, which is where the real pain lies.
> No need for this kind of bookkeeping since the dominator opts happen
> before we build those links.
>
> jeff
>