[LLVMdev] Proposal for new Legalization framework

[Krzysztof Parzyszek]

On 4/25/2013 9:58 AM, Dan Gohman wrote:
>
> It really needs some stiff refactoring. And optimization. And better
> documentation. And better developer tools. But with all the other
> problems, the zeitgeist has turned against it.

I appreciate the detailed explanation.  Thanks!

Speaking of the .td files and the table-driven instruction 
selection---yes, I think that there is a tendency for the .td files to 
become unreadable, and the selection process (divided into the custom 
and automatic lowering) to be somewhat harder to track and debug than 
the other passes.  In the long term, are there ideas to replace the 
tablegen with another tool?  The reason I ask is that I plan to develop 
a pattern-driven instruction combiner/simplifier for our Hexagon 
backend, that would work on the MI level, and I was planning to use 
tablegen and .td files to represent the patterns.  The only reason for 
that is that tablegen is already here, but I would be interested in any 
suggestions on how to represent MI patterns in ways that are (1) easy to 
edit/add/understand, and (2) not prohibitively complex/obscure for the 
code to work with (i.e. parse/match/substitute/etc.).

-Krzysztof

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, 
hosted by The Linux Foundation