llvm dev - May 2020 - Custom Instruction Cost Model to LLVM RISC-V Backend

Hi,

I’m working on a RISC-V architecture that has instruction costs different
from those in the default cost model. Is there an out-of-source way to
provide llc with custom cost model? Or, does this need a change in LLVM
backend?

Also, the cost model is not totally static. For example, loads from
0x1000-0x1ffc take 1 cycle, whereas loads from address > 0x80000000, take
10-100 cycles. Is it possible to model these costs in the scheduler?
Requiring programmer to define custom pointer types to mark pointers, to
one or the other address space, is okay. Is there a way to model different
costs for different pointer types?

Thank you,
Bandhav
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200525/264336ae/attachment.html>

Hi Bandhav,

While I'm unfamiliar with the details of codegen and cost modelling in
LLVM, it is possible to declare multiple address spaces for a target, mark
pointers as belonging to a specific address space, and cast between them.
To quote the language reference "For targets that support them, address
spaces may affect how optimizations are performed and/or what target
instructions are used to access the variable.", which sounds quite hopeful
for your usecase. If nothing else, you could probably model the loads from
different address spaces as being different instructions, even if they are
in the end the same instruction in the output machine code, and tag them
with different costs.

Hope this helps,
Henrik

On Tue, May 26, 2020 at 4:55 AM Bandhav Veluri via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Hi,
>
> I’m working on a RISC-V architecture that has instruction costs different
> from those in the default cost model. Is there an out-of-source way to
> provide llc with custom cost model? Or, does this need a change in LLVM
> backend?
>
> Also, the cost model is not totally static. For example, loads from
> 0x1000-0x1ffc take 1 cycle, whereas loads from address > 0x80000000,
take
> 10-100 cycles. Is it possible to model these costs in the scheduler?
> Requiring programmer to define custom pointer types to mark pointers, to
> one or the other address space, is okay. Is there a way to model different
> costs for different pointer types?
>
> Thank you,
> Bandhav
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200526/0dddd516/attachment.html>

Thanks a lot Henrik!

I figured following would mark a pointer to a specific address space:

#define __myaddrspace __attribute__((address_space(1)))

__myaddrspace int* data;

And, I was able to verify loads being annotated to be from addrspace 1 in
the generated IR. Would this work for automatic variables as well?

In regards to using this in the backend, do I have to just modify the
source, or is it possible to load a library at runtime, like we load a
frontend pass?

On Tue, May 26, 2020 at 12:37 PM Henrik Olsson <hnrklssn at gmail.com>
wrote:
> Hi Bandhav,
>
> While I'm unfamiliar with the details of codegen and cost modelling in
> LLVM, it is possible to declare multiple address spaces for a target, mark
> pointers as belonging to a specific address space, and cast between them.
> To quote the language reference "For targets that support them,
address
> spaces may affect how optimizations are performed and/or what target
> instructions are used to access the variable.", which sounds quite
hopeful
> for your usecase. If nothing else, you could probably model the loads from
> different address spaces as being different instructions, even if they are
> in the end the same instruction in the output machine code, and tag them
> with different costs.
>
> Hope this helps,
> Henrik
>
> On Tue, May 26, 2020 at 4:55 AM Bandhav Veluri via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> Hi,
>>
>> I’m working on a RISC-V architecture that has instruction costs
different
>> from those in the default cost model. Is there an out-of-source way to
>> provide llc with custom cost model? Or, does this need a change in LLVM
>> backend?
>>
>> Also, the cost model is not totally static. For example, loads from
>> 0x1000-0x1ffc take 1 cycle, whereas loads from address > 0x80000000,
take
>> 10-100 cycles. Is it possible to model these costs in the scheduler?
>> Requiring programmer to define custom pointer types to mark pointers,
to
>> one or the other address space, is okay. Is there a way to model
different
>> costs for different pointer types?
>>
>> Thank you,
>> Bandhav
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200527/6190da8f/attachment.html>