[llvm-dev] Struggling to define a simple instruction

[Rail Shafigulin via llvm-dev]

Hi Sam.

Thanks for the reply. This is not exactly what I was looking for.

Say I have a loop that looks like this:

sum = 0
for (i = 0; i < 4; i++)
  sum += (a[i] + b[i])

I would like a compiler to do the following:

vector.load vector.r0, addressOfA
vector.load vector.r1, addressOfB
vector.add vector.r0, vector.r0, vector.r1
sumreduce gpr.r0, vector.r0

Another scenario is adding all the elements of the array (this is probably
a better example)
sum = 0
for (i = i; i < 4; i++)
  sum += a[i]

would be converted to
vector.load vector.r0, addressOfA
sumreduce gpr.r0, vector.r0


If you can assist with this, I would greatly appreciate it.


On Tue, May 17, 2016 at 2:57 AM, Sam Parker <Sam.Parker at arm.com> wrote:

> Hi Rail,
>
>
> If you want to reduce the elements of a vector to a scalar sum, isn't the
> source operand a vector already? If this is the case, you'll want to
> extract each element and then use the three adds, something like:
>
>
> (add (add (add (extractelt VReg, 0), (extractelt VReg, 1)), (extractelt
> VReg, 2)), (extractelt VReg, 3))
>
>
> cheers,
>
> sam
>
>
> Sam Parker
>
> Graduate Engineer, Compilation Tools
>
> Development Solutions Group
> ------------------------------
> *From:* llvm-dev <llvm-dev-bounces at lists.llvm.org> on behalf of Rail
> Shafigulin via llvm-dev <llvm-dev at lists.llvm.org>
> *Sent:* 16 May 2016 22:56:25
> *To:* llvm-dev
> *Subject:* [llvm-dev] Struggling to define a simple instruction
>
> I'm trying to figure out how I can define and make LLVM target a simple
> instruction.
>
> The instruction is very simple, it adds all of the elements of the vector
> register. Lets' call it sum_reduce.
>
> Defining a vector add was simple. All I had to do is to simply change the
> register type used by a regular add instruction and that was it. However
> I'm really struggling to define this new instruction.
>
> Initially I thought that I would replace a chain of 3 adds with
> insert_element instructions, but that defeats the purpose of sum_reduce,
> since it would mean that I have to insert operators 4 times into a vector
> register and then add everything. This turns it into more instructions than
> the original chain of 3 adds.
>
> I've also looked at X86 HADD instructions but also have difficulty
> understanding how to implement something similar to it in my target.
> ,
> I've annoyed enough people on the LLVM IRC channel. And they have been
> really helpful, unfortunately I feel that if I ask more they'll stop
> answering completely :D
>
> At this point I wouldn't mind some handholding.  I can provide more code
> if needed
>
> Any help on the issue is really appreciated.
>
> --
> Rail Shafigulin
> Software Engineer
> Esencia Technologies
> IMPORTANT NOTICE: The contents of this email and any attachments are
> confidential and may also be privileged. If you are not the intended
> recipient, please notify the sender immediately and do not disclose the
> contents to any other person, use it for any purpose, or store or copy the
> information in any medium. Thank you.
>



-- 
Rail Shafigulin
Software Engineer
Esencia Technologies
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160517/efbfd8bb/attachment.html>