[LLVMdev] Question about per-operand machine model

[Andrew Trick]

On Feb 19, 2014, at 1:54 PM, jingu <jingu at codeplay.com> wrote:

> Hi Andy,
> 
> I am trying to schedule and packetize instructions for VLIW at post-RA
> stage or final codegen stage, where code transformations are not allowed
> any more, because hardware can not resolve resource conflict. There is a
> simple example as following:
> 
> ADD dest_reg1, src_reg1, src_reg2 (functional unit : ALU)
> STORE dest_reg2, mem (functional unit: LOAD_STORE)
> 
> These instructions can be genally packetized together because there is
> no dependency among operands and they use different functional unit. But
> we have one more restricton. The restriction is that some of
> instructions can not access to same register file at the same cycle. In
> other words, if 'src_reg1' of ADD instruction uses register file 'A' and
> 'dest_reg2' of STORE instruction uses same register file at the same
> cycle, it causes resource conflict and can not be executed on same
> cycle. This restriction depends on instruction type. I tried to consider
> each register file as a resource unit which is consumed by each operand.
> While scheduling instructions per cycle, used register file is recorded
> on state per cycle to check the conflict. In our heristic, it depends on
> operand's latency to record this resource on specific cycle's state. so
> I have tried to find a way to get latency and resource with each
> operand. If it is not possible to support this feature with per-operand
> resource model, as you suggested, I will try to make our own state
> machine or other scheduling constraint logic. I am newbee with
> scheduler. If you have any kinds of comment or feel something worng,
> please let me know. It will be really helpful.

It sounds like the register file is static and does not depend on register allocation. In this case, what you tried makes sense but is really not supported. The machine model tables are designed to be efficient for the common case, and per-operand resources don’t really make sense most of the time.

It sounds like you want to model the pipeline stage at which a resource is used. To do that with the per-operand machine model (misnomer), I think we need a ResourceDelay vector in addition to ResourceCycles, which we could easily add.

However, overall, I think you’re target is interesting enough that you may be better off augmenting the standard machine model with your own model. Your scheduler plugin could keep your own tables or state machine to model the constraints.

If you want to be clever, you could write tablegen code to build your model up from the SchedRead/Write definitions that are part of the standard model. You could add extra fields specific to your model.

Were you previously using the old instruction itineraries, and now moving to the new model?

-Andy

> 
> Thanks for your kind response,
> JinGu Kang
> 
> On 2014-02-20 오전 2:27, Andrew Trick wrote:
>> Hi JinGu,
>> 
>> We currently have the ResourceCycles list to indicate the number of cpu cycles during which a resource is reserved. We could simply add a ResourceDelay with similar grammar. The MachineScheduler could be taught to keep track of the first and last time that a resource is reserved.
>> 
>> Note that the MachineScheduler will work with the instruction itineraries if you choose to implement them. That’s the only way to get a full reservation table without customizing the scheduler. You can plugin your own state machine or other scheduling constraint logic. You may want to do this if you have very complicated constraints.
>> 
>> Can you provide an example of the most complicated instruction resources that you need to model?
>> 
>> -Andy
>> 
>> On Feb 19, 2014, at 4:57 AM, JinGu Kang <jingu at codeplay.com> wrote:
>> 
>>> Hi Andy,
>>> 
>>> I am sorry to misunderstand 'ReadAdvance' code. In order to support
>>> resource per operand, I feel we need more table and function. If
>>> possbile, I would like to listen to your opinion whether this feature is
>>> useful or not. As I mentioned on previous e-mail, it will be useful to
>>> access the latency and the resource per operand while checking resource
>>> conflict per cycle.
>>> 
>>> Thanks,
>>> JinGu Kang
>>> 
>>> On 18/02/14 23:09, jingu wrote:
>>>>> Resources and latency are not tied. An instruction is mapped to a
>>>>> scheduling class. A scheduling class is mapped to a set of resources
>>>>> and a per-operand list of latencies.
>>>> Thanks for your kind explanation.
>>>> 
>>>> Our heuristic algorithm have needed the latency and the resource per
>>>> operand to check resource conflicts per cycle. In order to support
>>>> this with LLVM, I expected a per-operand list of resources like
>>>> latencies with a scheduling class.
>>>> 
>>>> Can I ask you something to modify on tablegen? I think that the
>>>> 'WriteResourceID' field of 'MCWriteLatencyEntry' is for identifying
>>>> the WriteResources of each defintion as commented on code. As you
>>>> know, tablegen sets the 'WriteResourceID' field of
>>>> 'MCWriteLatencyEntry' with 'WriteID' when the 'Write' of defition is
>>>> referenced by a 'ReadAdvance'. If we always set this field with
>>>> 'WriteID', it causes problem? I can see that 'ReadAdvance' only uses
>>>> the 'WriteResourceID' field of 'MCWriteLatencyEntry' in
>>>> 'computeOperandLatency' function. I think the pair of latency and
>>>> write resource for defintion will be useful to check conflicts of
>>>> resources. As reference, I have attached simple patch.
>>>> 
>>>> Thanks,
>>>> JinGu Kang
>>>> 
>