[llvm-dev] [VLIW Scheduler] Itineraries vs. per operand scheduling

[Andrew Trick via llvm-dev]

> On Feb 4, 2018, at 9:15 AM, Yatsina, Marina via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> Hi,
>  
> What is the best way to model a scheduler for a VLIW in-order architecture?
> I’ve looked at the Hexagon and R600 architectures and they are using itineraries. I wanted to understand the benefit in using itineraries over the per operand scheduling.
>  
> I also found this thread from almost 2 years ago:
> http://lists.llvm.org/pipermail/llvm-dev/2016-April/098763.html <http://lists.llvm.org/pipermail/llvm-dev/2016-April/098763.html>
>  
> At that time it seemed the itineraries are a better choice, but is it still the case?
> Also, in this thread Phil says:
> “Some of the constraints that can be found in in-order micro architectures cannot be expressed in the per-operand scheduling model”
> Does anybody have an example of such constraints that will be harder to model with per operand scheduling?
>  
> Thanks,
> Marina

Hi Marina,

Itineraries have stuck around mainly because no one was interested in porting old itineraries to a new machine description. At least initially, there also wasn't anyone interested in working on an in-order scheduling implementation based on the new machine model, even though the machine description itself was designed for in-order scheduling. There are a handful of targets now using the new model with MicroOpBufferSize = 0/1 (in-order mode). It's hard for me to say how widely supported it is, because so many backends using the MachineScheduler are out-of-tree.

The per-operand model should be much more compile time efficient, but that's often not a concern. It can be easier now to incrementally bootstrap new machine descriptions based on similar machines with the same architecture. The subtarget details no longer need to be tied to the TableGen instruction descriptions. The new model lets you define the structure of the subtarget scheduling description, so it can be self-documenting. The new model lets you use a plethora of tablegen tricks to handle all the variations on vector operations. I believe it's much more straightfoward to specify special constraints, like bypassing. It's fairly easy to extend certain opcodes with predicates (callbacks) that depend on immediate values or any other arbtrary info.

See Javed and Florian's tutorial on the "new" machine model:
https://www.youtube.com/watch?v=brpomKUynEA

Itineraries are not strictly more powerful. The original motivation for the new machine description is that there were important issue constraints that weren't being correctly modeled by the itineraries. The only advantage of itineraries that I'm aware of is that, for instructions that use multiple machine resources, you can control during which cycle those resources are used relative to each other (it specifies a full resource matrix for each op). The per-operand model makes the simplifying assumption that a given resource is simply used for N cycles. Normally, all instructions begin using a given resource on the same cycle relative to instruction issue, so that's not an innaccurate assumption. If you actually need to correct for that, and don't use itineraries then you need to make use of the the MachineScheduler's hooks to write a little custom scheduling code.

Whatever you do, there's no sense in using both kinds of machine descriptions. Support is only there for the purpose of bootstrapping using ARM A9, since that was the latest in-tree ARM cpu at the time. It's just useless complexity now.

-Andy

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180207/036bb809/attachment.html>