Re: basic block output order?

[Linus Torvalds <torvalds@osdl.org>]

On Tue, 12 Dec 2006, Christopher Li wrote:
>
> I am playing with the example.c. I am wondering why the parent
> basic blocks need to generate first?

They don't have to. But it generated nicer code, iirc, mainly because it 
did the storage allocation the natural way. In particular, if I recall 
correctly, it causes loops to have the storage for the _innermost_ loop to 
be done first.

Notes off the top of my head, without actually looking at the code: 
because when you hit a loop, the "parent" set is actually both the entry 
and the BB that ha the loopback, so you actually end up going to the 
loopback thing, which goes to _its_ parents, etc etc, until you actually 
get back to the _top_ of the loop (and now the "generation" count triggers 
you to break the looping), so you end up doing the actual register 
allocation at tops of loops, but because you do this all recursively, and 
the inner loop will have _its_ parents point to the callback too, you 
generally tend to have started storage allocation at loop-tops.

HOWEVER. There's a reason the thing is called "example.c". The reason is 
simply that it's stupid, idiotic, and not meant to be taken seriously. I 
also ended up just hackign things around randomly to make it output 
something that looked half-way sane, _and_ I ended up changign it to use 
the "unssa" pass by Luc, _and_ I'm border-line psychotic when it comes to 
compilers anyway.

In other words, what I'm trying to say is that you shouldn't take anything 
I say too seriously, and that the "example.c" code wasn't really even 
meant to be serious. I always wanted somebody else to write the back-end, 
and held back as long as I could from writing example.c, and when I wrote 
it, it was more a case of desperately trying to find somebody interested 
in it, and having it as an example of how things _might_ work.

So as far as I am concerned, the whole "example.c" is just total 
throw-away code. Go wild with it.

		Linus