[LLVMdev] Make LoopBase inherit from "RegionBase"?

[Tobias Grosser]

On 01/15/10 03:03, ether wrote:
> hi,
Hi,

> about region class, i think provide some method to extract the top level
> loop from a given region is useful in polly(our polyhedral optimization
> framework), because the we are only doing optimization on the region
> with loop.

Hi ether,

I think this should be implemented as a RegionFilter, that checks if a
region contains a loop, and that can be asked for further information.
In general I do not think this kind of analysis belongs to a region, but
as you proposed some kind of filter could be applied. In the short term
the passes who need this information could get it on their own.

> and a long term consideration about "region pass":
> 
> if we want to integrate region analysis and optimization framework into
> llvm,  i think we can use an approach that similar to loop analysis and
> optimization: write a class "regionpass" inherit from "pass", and the
> corresponding pass manger "RegionPassManager". (a kind of function pass)
> if we follow this approach, we need to push the region pass manager into
> the llvm pass manager stack.
> the first question of this approach is, whats the relationship between
> "loop pass manager" and "region pass manager"?
> 
> way 1: make region pass manager below loop pass manager in the stack
> 
> pass manager stack:
> 
> bb pass manager <---top
> loop pass manager
> region pass manager
> function pass manager
> ... <---bottom
> 
> in this way the region hierarchy need to be reconstruct when a loop
> transform change it.
> 
> way 2: make region pass manager above loop pass manager in the stack
> 
> pass manager stack:
> 
> bb pass manager <---top
> region pass manager
> loop pass manager
> function pass manager
>     ... <---bottom
> 
> in this way the loop hierarchy need to be reconstruct when a region pass
> change it.
> 
> now we need to choose a way to minimize the loop reconstruction or
> region reconstruction. i think that the chance that a region transform
> affect the loop structure is smaller, so maybe way 2 is better.
This would need some thoughts. Ideal I think we would not order them, but if
a region changed, just reconstruct the loops that are in this region and
if a
loop changed just reconstruct the regions in this loop.

> at last, i have some idea about finding a interesting region: (maybe
> make the region analysis too complex)
> 
> we can introduce some thing like "region filter" that determine the
> property of a region, the region filter will like a "pass", which can
> run on an instruction at a time, a basic block at a time, or even a sub
> region at a time, then we can write a "filter manager" like "pass
> manager " to stream the filtering process, so that we can promote the
> the performance of the region finding process.
Yes, I like this idea.
So the basic design would be that we have some passes like:
Maximal Region regarding an Analysis/Filter
Minimal Region regarding an Analysis/Filter
All Regions regarding an Analysis/Filter

So a pass can ask the regionpass manager for a specific kind of regions.
It is than just invoked for regions, that fulfill this requirement.

Tobias