[LLVMdev] RFC: Pass Manager Redux

[John Criswell]

On 7/11/12 4:50 AM, Chandler Carruth wrote:
> Greetings folks!
>
> In working on a new optimization pass (splitting cold regions into 
> separate functions based on branch probabilities) I've run into some 
> limitations of the current pass manager infrastructure. After chatting 
> about this with Nick, it seems that there are some pretty systematic 
> weaknesses of the current design and implementation (but not with the 
> fundamental concepts or behavior).
>
> Current issues:
>
> [snip]
>
> - What else did I miss?

Analysis groups.

There are two limitations on analysis groups:

1) For some reason, setting one up and adding passes to it doesn't seem 
to have worked quite as smoothly as it should.  I don't remember the 
exact details, but in the past, I've had to struggle to create Analysis 
Groups and make them work properly.

2) Rerunning analysis group passes after invalidation.  If you want to 
use an analysis other than the default analysis group, you have to 
manually specify where to use it each time in the pass pipeline. If your 
non-default pass gets invalidated, then the default is used by any other 
transform that requests an analysis from that group.

-- John T.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20120711/25b1f97c/attachment.html>