This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [stree] Committed: initial stree infrastructure, and stree version of enumerators


On Mar 9, 2004, at 2:17 AM, Mark Mitchell wrote:

Matt Austern wrote:

First installment of the stree work. We started with enumerators because they're simple, and also because in some environments (e.g. files that include Apple's Carbon.h header) there are an awful lot of them. The goal of this work is to avoid having the compiler allocate memory for unused declarations, and, in any real program, the vast majority of Carbon.h enumerators are never used.

To avoid some of the communication issues that we've had with other improvements, I want to make clear that I think that y'all should post an informal design document with goals, implementation strategy, justification of that strategy, etc. before merging. I'd like to see a write-up, not just a patch. In particular, I bet I will things to say (positive or negative!) about this approach, but I'm not going to try to examine your preliminary patch to make a determination about that.


My initial instinct is that this is not the most high-value target, and that the resulting additional messiness in the type system is going to outweigh other wins. There are a lot of algorithmic mistakes in the compiler, and I'd rather we fix them, instead of complicating the infrastructure further to compensate for our other stupidity. :-) But, that's all somewhat in the nature of anecdotally-reinforced hypothesis, not science.

That's a good point. Ideally Geoff and I should have collaborated on that document before he left on vacation, but I'll do what I can on my own.


It'll have to be an *informal* design document, because some of the details of the stree patch I committed yesterday are going to change. I think there's a very good chance, for example, that this whole page system for memory management is premature optimization. But the basic goal and strategy is fairly simple.

I'm definitely sensitive to the problem of gcc "improvements" that make the compiler even more complicated and harder to understand, and I certainly agree that there are algorithmic improvements to be had as well. But I also think the stree approach is promising enough that we ought at least to explore it. We need to make major speedups, not just avoid further slowdowns.

--Matt


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]