[PATCH 1/5] refactor LiveInterval: introduce new LiveRanges class

[Matthias Braun]

The name LiveRanges has been changed in the second iteration of the patches it’s indeed called LiveRange in singular form now (which is indeed a better name). This is the structure after applying this patchset:

* A LiveRange object contains a list of LiveRange::Segment objects which represent intervals
* LiveIntervall is simply a subclass of LiveRange adding a field for the register number and spill weight
* Several algorithms that operated on LiveInterval but weren’t really bound to a specific register/machine instructions now work on the LiveRange base class.
* Not in this patch series, but I’m working on this: Introduce a new LiveIntervall::SubRange class which is a subclass of LiveRange with a LaneMask added, managing liveness of a “subregister”/a group of lanes.

Greetings
	Matthias

On Sep 6, 2013, at 4:16 PM, Jakob Stoklund Olesen <stoklund at 2pi.dk> wrote:

> 
> On Sep 6, 2013, at 3:36 PM, Andrew Trick <atrick at apple.com> wrote:
> 
>> 
>> On Aug 29, 2013, at 6:07 PM, Matthias Braun <matze at braunis.de> wrote:
>> 
>>> 
>>> LiveRanges has logic to manage a set list LiveRange instances
>>> and a list of value numbers which previously was in LiveInterval,
>>> without having details like spill weight or a fixed register number.
>>> LiveInterval is now a direct subclass of LiveRanges and simply adds the
>>> spill weight and the register number.
>>> ---
>>> include/llvm/CodeGen/LiveInterval.h  | 143 +++++++++++++++++++++--------------
>>> lib/CodeGen/LiveInterval.cpp         | 128 +++++++++++++++----------------
>>> lib/CodeGen/LiveIntervalAnalysis.cpp |  20 ++---
>>> 3 files changed, 156 insertions(+), 135 deletions(-)
>>> 
>>> <0001-refactor-LiveInterval-introduce-new-LiveRanges-class.patch>_______________________________________________
>> 
>> I don’t want to hold up patches with philosophical arguments—and it’s not your fault that some terminology is already misused.
> 
> I think we should use this opportunity to get the names right. We’re causing massive churn anyway.
> 
>> This is just something to think about:
>> 
>> In compiler literature a LiveRange is an allocatable entity as seen by the register allocator.
>> 
>> By definition, a range is a disjoint set of intervals where each interval represents all values between two end points. So in the context of a compiler, interfence is computed by checking intersection of intervals.
> 
> It seems that one of these definitions will have to give since we want to represent the liveness of a virtual register as a linked list of {mask + disjoint set of intervals}.
> 
> /jakob
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130906/9a013b2e/attachment.html>