[llvm-commits] [LLVMdev] [PATCH] Refactoring the DFA generator

Anshuman Dasgupta adasgupt at codeaurora.org
Fri Aug 24 08:01:29 PDT 2012

Hi Ivan,

> I missed last 2 commits made by Alexey. Following his advices, I 
> updated the patch. It should be ok now.
> Thanks Anshu!
> I've recently added more functional units to our Schedule.td and the 
> generation time became painfully long. In fact, the main problem was 
> in writeTableAndAPI(). I propose another patch to fix it:
> - Fixed memory leaks.
> - Transitions are attached to states.
> I've regenerated the DFA table of Hexagon and everything is ok.
> Please review it.

I had a few comments about the design change that you're introducing 
with the patch:

This patch adds multiple points of control for adding a transition: 
there's now an addTransition() for DFA and another addTransition() for 
State that populate different data structures. We shouldn't need both. I 
am okay with transitions being folded into the State class if it results 
in significant speedup in DFA generation. But then please remove the 
Transition mechanism for the DFA class (stateTransitions).

For transitions being folded into State:
 > std::map<unsigned, Transition *> inputToTrans;

We don't need a map to Transition* here; we can directly map from input 
to stateNum. You should be able to use a LLVM data structure.


More information about the llvm-commits mailing list