Adding support for SIMD datatypes and basic vector operations about mir HOT 4 CLOSED

Thank you for interesting in MIR.

I was asked this question several times.

To be honest right now I'd like to keep MIR away from implementing vector insns. It is not important feature for implementing lighweight tier1 JITs.

It is a very target dependent feature. Besides arithmetic insns you need insns to pack and unpack vectors, other necessary insns. Also why only fp insns? It would be logical to add integer vector insns too. Also you need new vector types, a lot of target dependent machine code to implement the vector insns (call ABI could be complicated even more). For me next logical step would be implementing autovectorization in MIR. It is so complicated area requiring a lot of expertise and efforts to implement huge code, a whole can of worms. So adding even simple vector insns would considerably complicate MIR.

I am not saying that MIR will never support vector insns. Simply I'd like to try MIR in Ruby JIT implementation and I don't see how vector insns could help this goal. If I see that MIR could be successful as tier2 compiler I probably will think hard about adding vector insns and which ones.

from mir.

I agree with @vnmakarov here. We have a similar use case with Faust and we just added a backend to use MIR, see #54.

The thing is that MIR seems perfect to deploy reasonably fast code using a small an easy to deploy piece of code. It you really want the fastest code, then you'll have to use the LLVM path.

from mir.

Hi, I think MIR should avoid the complexity and bloat from vector insns because there is no need for MIR to be like LLVM. The huge strength of MIR is small size and reasonable performance. I would hate to see feature creep in MIR that then makes it lose its special asset. No point in making another LLVM.

from mir.

Ok. Thank you for your answers. Anyway, I will try to add some basic SIMD stuff in a local branch to see if is feasible.

My purpose of using SIMD in MIR is not to create a super-optimizin compiler, but it is including the SIMD datatypes and operations in the IR itself in order to avoid external add-ons to a language as intrinsics are used in C/C++.

from mir.