llvm dev - Feb 2017 - Pre-RA scheduler does not generate NOPs when getHazardType() returns NoopHazard

Hello.
     I am new to the schedulers implemented in the back end of LLVM.
     I am trying to handle data hazards in my simple processor, with
instructions that
execute in 1 cycle.

     I have tried the standard post-RA scheduler, implemented in 
lib/CodeGen/PostRASchedulerList.cpp, (with a ScoreboardHazardRecognizer), but I
have some
issues with some consecutive instructions that are separated in the final ASM
listing and
they should not.

     I also tried the standard pre-RA scheduler, implemented in 
lib/CodeGen/ScheduleDAGRRList.cpp, (with a ScoreboardHazardRecognizer), but even
if I
report that I have a data hazard in getHazardType() and return NoopHazard, the
scheduler
will not insert a NOP instruction in the respective place where I report the
hazard . For
me it is more convenient, in principle, to use the pre-RA scheduler, but it is
not
generating NOPs.
     Could you please tell me how can I generate NOPs with a pre-RA scheduler
when I am
already informing it that at the current place we have a NoopHazard.

  Thank you,
     Alex

   PS: Also, could you please tell me if with the standard pre-RA scheduler with
ScoreboardHazardRecognizer I could also employ other useful program instructions
when
dealing with hazards, instead of using NOPs.

Hello.
     I am not sure if this is the correct answer to the question I put in the
previous
email, but please take a look at what I wrote in another thread at 
http://lists.llvm.org/pipermail/llvm-dev/2017-March/111293.html .
     Basically, I describe there I was more successful to use PreEmitNoops(),
instead of
the getHazardType() method which, when I return NoopHazard, it does NOT always
generate a
NOP - in my case it works for Store, but not for Load MachineInstructions.

     Does the ScoreboardHazardRecognizer class generate (or not) a NOP
instruction, when
instructed to in getHazardType(), depending on the type of the instruction
analyzed?

   Thank you,
     Alex


On 2/12/2017 4:10 AM, Alex Susu wrote:
>   Hello.
>     I am new to the schedulers implemented in the back end of LLVM.
>     I am trying to handle data hazards in my simple processor, with
instructions that
> execute in 1 cycle.
>
>     I have tried the standard post-RA scheduler, implemented in
> lib/CodeGen/PostRASchedulerList.cpp, (with a ScoreboardHazardRecognizer),
but I have some
> issues with some consecutive instructions that are separated in the final
ASM listing and
> they should not.
>
>     I also tried the standard pre-RA scheduler, implemented in
> lib/CodeGen/ScheduleDAGRRList.cpp, (with a ScoreboardHazardRecognizer), but
even if I
> report that I have a data hazard in getHazardType() and return NoopHazard,
the scheduler
> will not insert a NOP instruction in the respective place where I report
the hazard . For
> me it is more convenient, in principle, to use the pre-RA scheduler, but it
is not
> generating NOPs.
>     Could you please tell me how can I generate NOPs with a pre-RA
scheduler when I am
> already informing it that at the current place we have a NoopHazard.
>
>  Thank you,
>     Alex
>
>   PS: Also, could you please tell me if with the standard pre-RA scheduler
with
> ScoreboardHazardRecognizer I could also employ other useful program
instructions when
> dealing with hazards, instead of using NOPs.

On 03/21/2017 08:29 AM, Alex Susu via llvm-dev wrote:
>   Hello.
>     I am not sure if this is the correct answer to the question I put 
> in the previous email, but please take a look at what I wrote in 
> another thread at 
> http://lists.llvm.org/pipermail/llvm-dev/2017-March/111293.html .
>     Basically, I describe there I was more successful to use 
> PreEmitNoops(), instead of the getHazardType() method which, when I 
> return NoopHazard, it does NOT always generate a NOP - in my case it 
> works for Store, but not for Load MachineInstructions.
>
>     Does the ScoreboardHazardRecognizer class generate (or not) a NOP 
> instruction, when instructed to in getHazardType(), depending on the 
> type of the instruction analyzed?
What do you mean? The ScoreboardHazardRecognizer does not generate NOP 
hazards. You need to override it to do that. The post-RA scheduler 
should then generate the NOPs (by calling your insertNoop).

  -Hal
>
>   Thank you,
>     Alex
>
>
> On 2/12/2017 4:10 AM, Alex Susu wrote:
>>   Hello.
>>     I am new to the schedulers implemented in the back end of LLVM.
>>     I am trying to handle data hazards in my simple processor, with 
>> instructions that
>> execute in 1 cycle.
>>
>>     I have tried the standard post-RA scheduler, implemented in
>> lib/CodeGen/PostRASchedulerList.cpp, (with a 
>> ScoreboardHazardRecognizer), but I have some
>> issues with some consecutive instructions that are separated in the 
>> final ASM listing and
>> they should not.
>>
>>     I also tried the standard pre-RA scheduler, implemented in
>> lib/CodeGen/ScheduleDAGRRList.cpp, (with a 
>> ScoreboardHazardRecognizer), but even if I
>> report that I have a data hazard in getHazardType() and return 
>> NoopHazard, the scheduler
>> will not insert a NOP instruction in the respective place where I 
>> report the hazard . For
>> me it is more convenient, in principle, to use the pre-RA scheduler, 
>> but it is not
>> generating NOPs.
>>     Could you please tell me how can I generate NOPs with a pre-RA 
>> scheduler when I am
>> already informing it that at the current place we have a NoopHazard.
>>
>>  Thank you,
>>     Alex
>>
>>   PS: Also, could you please tell me if with the standard pre-RA 
>> scheduler with
>> ScoreboardHazardRecognizer I could also employ other useful program 
>> instructions when
>> dealing with hazards, instead of using NOPs.
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory