[llvm-commits] [PATCH] Memory Dependence Analysis: differenciate "unknown" from "no dependency in current function"
Xiaoyi.Guo at amd.com
Wed Oct 12 15:17:55 CDT 2011
Thanks a lot for your feedback. I'm working on re-writing it according to your suggestion.
From: Eli Friedman [mailto:eli.friedman at gmail.com]
Sent: Tuesday, October 11, 2011 5:27 PM
To: Guo, Xiaoyi
Cc: llvm-commits at cs.uiuc.edu
Subject: Re: [llvm-commits] [PATCH] Memory Dependence Analysis: differenciate "unknown" from "no dependency in current function"
On Tue, Oct 11, 2011 at 4:57 PM, Guo, Xiaoyi <Xiaoyi.Guo at amd.com> wrote:
> If the memory dependence analysis did not find any dependence for a pointer in the current function, it currently returns "Unknown" as the result. However, in the optimization that I'm doing, I need to be able to tell this case apart from a general "unknown" case. i.e., I need to know that the pointer does not have any dependencies in the current function. The patch attached is to achieve this.
> Please help to review and commit if acceptable.
The interface to MemDepResult::getNonFuncLocal makes no sense: as far as I can tell, the Instruction* which is passed in is useless except for the fact that it isn't null.
MemDepResult's internal representation is starting to get a bit crazy... with your changes, there are 6 states: Invalid, Def, Clobber, NonLocal, NonLocalFunc, and Unknown. Def and Clobber have an associated instruction, the other states do not. Granted, I sort of started the craziness by the way Unknown is represented, but this is starting to become a complete mess; maybe we can restore a bit of sanity by splitting things up so that the bottom part of the pair is "Def", "Clobber", or "Other", and encode the other states into upper part of "Other", or something like that?
More information about the llvm-commits